perf: implement fast Get for integral types

[TerrorJack]

This patch implements fast Get logic for integral types based on:

• Use a single load operation when loading with same endianness of the
  host, otherwise do a host load and a byteSwap. This avoids the
  overhead of multiple single-byte loads in the previous
  implementation.
• Use the unaligned Addr# load/store primops added since GHC 9.10 when
  available, otherwise do a plain peek. This ensures the GHC backends
  see the right AlignmentSpec at the Cmm level and can correctly emit
  unaligned load instructions.

There's no need for changing Put logic they're backed by FixedPrim
logic in Data.ByteString.Builder.Prim.Binary that already does
similar optimization.

Closes #215.

[Bodigrim]

(I'm not a maintainer here)

[Bodigrim]

Is it feasible to add a s390x job to CI? See https://github.com/haskell/bytestring/blob/master/.github/workflows/ci.yml#L121 for instance. Otherwise #if defined(WORDS_BIGENDIAN) tends to bit rot really quickly.

[TerrorJack]

that'll be an extra source of flakiness before https://gitlab.haskell.org/ghc/ghc/-/issues/25541 is sorted out

[konsumlamm]

I think it's a good idea to run CI on a big endian arch, but that can be done in a later PR.

[konsumlamm]

I tried to benchmark the use of the unaligned load primops, but I couldn't notice any difference. Do you have any idea why that might be?

This is my benchmark:

import Control.Monad

import Data.Binary.Get
import Data.Binary.Put

import Test.Tasty.Bench

main :: IO ()
main = defaultMain
    [ bench "" $ whnf (runGet getData) bs
    ]
  where
    n = 100000
    getData = fmap sum $ replicateM n $ do
        w8 <- getWord8
        w16 <- getWord16host
        w32 <- getWord32host
        w64 <- getWord64host
        pure $! fromIntegral w8 + fromIntegral w16 + fromIntegral w32 + w64
    bs = runPut $ replicateM_ n $ do
        putWord8 42
        putWord16host 12
        putWord32host 0xff00ff00
        putWord64host 0x0123456789abcdef