Relax bounds for aeson and optparse-applicative in test-suite

Split sections support

Solver: Enforce dependencies on executables (fixes #4781).

Add --enable-split-sections flag and pipe it through to the GHC backend. Note
that some of the implementation here could be made a bit more precise:
-split-sections and -split-objs are mutually exlusive yet the types don't
currently reflect this.

Fixes #4819.

Examples which used the ancient spec-version v1.2 have been updated to
use at least v1.8; and use CPP-Options instead of GHC-Options where
`cabal check` would have warned; avoid automatic flags for features;
also be explicit about the `build-type:` to avoid the implicit `Custom`
default (TODO: we should really change the default with a future
spec-version to `Simple`)

[skip ci]

Change order of hsLibraries

After using this for a while. It occured to me that due to linker ordering, and GHC just forwarding the libraries as specified in the `hsLibraries` field, we should put the `extraBundledLibs` at the end. As such the haskell library can reference the `extraBundledLibs`.

This still leave the question of recursively linked libs. But getting that linked right should rather be deal with in GHC, which optimally would use being and end groups for the libraries to link.

Add binary distribution download link to Readme

I've so often failed to find that link. Let's just add it to the readme.

Allow empty set of input modules.

This reverts commit 6766ef73.

Now also caches cabal-install deps.

Properly list 'extra-library-flavours' instead of 'virtual-modules' for `extraLibFlavours`.

Add `checkVersion` for added fields

Add `checkVersion` for asm-sources, cmm-source, extra-bundled-libaraires, extra-library-flavours, and virtual-modules.

As pointed out by @hvr, #4857, #4875 did not contain the necessary "check" logic. This PR tries to address this shortcoming.

Fixes #4890

Extension: Add HexFloatLiterals

The pre-new-build code path sets the solver parameter 'solveExecutables' to
false, which causes the solver to ignore all build-tools and build-tool-depends
dependencies.  This commit adds a test for ignoring each of the two types of
dependencies.

This commit adds two checks to the validation phase of the solver:

1. It checks that each newly chosen package instance contains all executables
   that are required from that package so far.

2. It checks that each new build tool dependency that refers to a previously
   chosen package can be satisfied by the executables in that package.

This commit also fixes a TODO related to solver log messages. Previously, it was
possible for the log to associate an incorrect executable name with a
dependency.

This commit changes the field of type 'IsExe' in the 'Dep' data type to
type 'Maybe UnqualComponentName'.  It also adds the executable name to error
messages that previously just contained "(exe)".

This commit splits the 'Conflicting' constructor of 'FailReason' into five
different constructors, for different types of dependency conflicts.  This
change makes it easier to print different error messages for different
conflicts, though it doesn't change any of the error messages yet.

The refactoring caused a small change to the error message for conflicts caused
by self-dependencies.

Fixing the TODO caused the solver to report conflicting constraints in the
opposite order in some cases, but the change didn't seem to have a negative
impact on the clarity of the error messages.

Here is an example of the change from running
"cabal install --dry-run nerf --index-state=2017-11-04T18:36:30Z -v3":

 [_19] trying: template-haskell-2.11.1.0/installed-2.1... (dependency of tagged)
 [_20] next goal: pretty (dependency of template-haskell)
-[_20] rejecting: pretty-1.1.3.3/installed-1.1... (conflict: pretty => deepseq==1.4.2.0/installed-1.4..., text => deepseq>=1.1.0.0 && <1.4)
+[_20] rejecting: pretty-1.1.3.3/installed-1.1... (conflict: text => deepseq>=1.1.0.0 && <1.4, pretty => deepseq==1.4.2.0/installed-1.4...)
 [_20] rejecting: pretty-1.1.3.5, pretty-1.1.3.4, pretty-1.1.3.3, pretty-1.1.3.2, pretty-1.1.3.1, pretty-1.1.2.1, pretty-1.1.2.0, pretty-1.1.1.3, pretty-1.1.1.2, pretty-1.1.1.1, pretty-1.1.1.0, pretty-1.1.0.0, pretty-1.0.1.2, pretty-1.0.1.1, pretty-1.0.1.0, pretty-1.0.0.0 (conflict: template-haskell => pretty==1.1.3.3/installed-1.1...)
 [_20] fail (backjumping, conflict set: pretty, template-haskell, text)

Adds `virtual-modules` to handle cases like `GHC.Prim` from `ghc-prim`.

GHC's `ghc-prim` package has an `exposed-module` for which there exists no source code. As such it can't be built and has no interface file.  With the goal of getting rid of hacks in GHC `ghc-prim`s cabal file, which does the following:

```
flag include-ghc-prim
    Description: Include GHC.Prim in exposed-modules
    default: False
[...]
    if flag(include-ghc-prim)
        exposed-modules: GHC.Prim
```

and has this explaination in the build system:

```
# ----------------------------------------
# Special magic for the ghc-prim package

# We want the ghc-prim package to include the GHC.Prim module when it
# is registered, but not when it is built, because GHC.Prim is not a
# real source module, it is built-in to GHC.  The old build system did
# this using Setup.hs, but we can't do that here, so we have a flag to
# enable GHC.Prim in the .cabal file (so that the ghc-prim package
# remains compatible with the old build system for the time being).
# GHC.Prim module in the ghc-prim package with a flag:
#
libraries/ghc-prim_CONFIGURE_OPTS += --flag=include-ghc-prim

# And then we strip it out again before building the package:
define libraries/ghc-prim_PACKAGE_MAGIC
libraries/ghc-prim_dist-install_MODULES := $$(filter-out GHC.Prim,$$(libraries/ghc-prim_dist-install_MODULES))
endef
```

The addition of the `virtual-modules` has thus the effect of a module that is assumed to exist
but not tried to be built or listed in any of the module files.  It is however part of the
`exposed-modules` when registered into the package database

Disable tests and benchs by default on new-install

[ci skip]

And error if they are explicitly enabled, so we never try to install
remote packages with tests and/or benchs.

Fixes #4855

[ci skip]