GHC merge request !963 improved warnings in the presence of
COMPLETE annotations. This allows the removal of the Fun pattern
from the complete set.

This patch accounts for the resulting changes in pattern match warnings.

Bump version to 0.8.9.0

Changes for GHC 9.2.1

Otherwise GHC 9.2 will compile the package as GHC2021, which will break
due to a Safe Haskell error. See GHC #19631.

Bring in changes GHC 9.2 is using

Remove instance for Data.Semigroup.Option for GHC >= 9.2

Use unsafe shifts from base rather than hand-rolling

See
https://hackage.haskell.org/package/base-4.5.0.0/docs/Data-Bits.html#v:unsafeShiftL.
in binary's minimum supportedbase version.

Beyond reducing duplicated code, this will avoid CPP in dealing with the
changing primop situation for GHC 9.2.

This is no longer needed with the next commit solving the problem
without CPP.

This reverts commit b2244101.

Corresponding GHC MR:
ghc/ghc!4945

Corresponding GHC MR:
ghc/ghc!4945

See
https://hackage.haskell.org/package/base-4.5.0.0/docs/Data-Bits.html#v:unsafeShiftL.
in binary's minimum supportedbase version.

Beyond reducing duplicated code, this will avoid CPP in dealing with the
changing primop situation for GHC 9.2.

Reading the file with 12000+ examples is slow as it's deserialized with
'read'.

Use bases built-in float/word32 and double/word64 casts

Travis keeps breaking due to its very fragile build script.
Switching to an approach which I hope will be more robust.

Only use recent version of cabal when building for all GHC versions.
This means we no longer test older cabal install binaries.

cabal-install is unable to come up with a build plan for the tests and
benchmarks. The problem seems to be that building binary's
tests/benchmarks requires having binary, which is what we're trying to
build. Cabal doesn't seem to separate the lib/tests/benchmarks into
distinct build units, so it fails to come up with a plan.

To work around this, when building on travis, we rename the binary
library. With a new name, cabal can find a plan and run the tests.

Remove GHC 7.4.2 and 7.6.3 from the testing matrix.
They fail with this new build setup for a reason I haven't investigated.

Remove a redundant catch-all clause

Handle {Int,Word}32Rep

The clause will be rendered redundant by [GHC MR 963](ghc/ghc!963).

Remove 7.4.1 as it's broken upstream.

Prepare 0.8.7.0

Haddock markup fixes

  * Use the standard `@since` notation (since it plays well with
    re-exports in other packages)
  * Use `@ ... @` for monospacing things that aren't identifiers
  * Qualify some linked identifiers (so that they actually link)

'Binary (NonEmpty a)' should 'fail' on empty lists

Co-Authored-By: harpocrates <alec.theriault@gmail.com>

Compatibility with future ghc-8.8/base-4.13

Typofix