Make it a looped KeyCommand inside a single call to runInputCmdT,
rather than calling runInputCmdT once for each character.

This hopefully makes it closer to "real" results when entering
multiple keys (e.g., holding them down).

This works with `stack` and `cabal new-test`.  The latter
uses the `build-tool-depends` field to locate the binary.

Also:
- Remove the reference to the nonexistent main file on Windows.
- Make the test binary return a nonzero exit code when it fails.

Still TODO: test in CI.

This change is motivated to make it easier to run the unit tests:

defined two executable targets in the cabal file
added run_unit_tests.sh script
one test is failing independent of the changes:

```
Failure in: 0:interaction:2:tab completion:0
tests/RunTTY.hs:106
expected: ["dummy-\206\188\206\177\207\131/\r\nbar \207\130\206\181\207\129\207\132\r\n0:dummy-\206\188\206\177\207\131/"]
but got: ["dummy-\206\188\a\a"]
```

Fixes #74 as well enough as Haskeline currently can.  It returns
a `Bool` rather than `Maybe Key` since Haskeline doesn't expose its
`Key` type.  But this should be enough for the "press any key to continue"
functionality described in the feature request.

Fixes #72.

Fixes #87.

Similar to `KillLine`, add two new "keys" `ReverseSearchHistory` and `ForwardSearchHistory` which bind to the incremental history search functionality.

Original patch by @boris-stepanov (#90).

Fixes #3.

Fixes #21, which noted that the `LICENSE` file was inconsistent with
`License: BSD3` in the .cabal` file.

Removes `System.Console.Haskeline.MonadException`; now
we use class constraints from `Control.Monad.Catch`.

- Use Travis's build-in apt support
- Use `language: c` explicitly
- Remove unnecessary `cabal` invocations
- Remove the separate `before_install` step

Also:
- Add CI testing up to ghc-8.6.1
- Don't install explicit Cabal versions, now that we don't have
  a custom Setup script.

Haskeline doesn't have a custom Setup script, so that step isn't
providing any value.

Helps fix `stack upload`.

Bump upper bound on containers

Bump stm upper bound

Bump upper bound on unix to < 2.9

See https://ghc.haskell.org/trac/ghc/ticket/15042.

Bump upper bound on base to < 4.13

See https://ghc.haskell.org/trac/ghc/ticket/15018.

Also bump to 0.7.4.2.

Unfortunately trac.haskell.org doesn't exist anymore.

Import Control.Concurrent.STM in Win32 backend

Move away from Chan to STM

The isEmptyChan broken has been long known to be broken

Make sections Markdown correct in README.md