Use unboxed Maybe for map lookups on GHC

[meooow25]

Trying this out for #801.

[meooow25]

The benchmarks results are interesting.

Map:

Name              Time - - - - - - - -    Allocated - - - - -
                       A       B     %         A       B     %
lookup absent     474 μs  502 μs   +5%    145 B   317 B   +118%
lookup present    469 μs  459 μs   -2%    128 KB  329 B   -99%

Great reduction in allocations, but no effect on time saved. I expect GHC is just very fast with short-lived allocations.

IntMap:

Name             Time - - - - - - - -    Allocated - - - - -
                      A       B     %         A       B     %
lookup_half      279 μs  296 μs   +6%    148 B   147 B    +0%
lookup_hits      479 μs  522 μs   +8%    1.0 KB  335 B   -67%
lookup_misses    472 μs  499 μs   +5%    309 B   305 B    -1%
lookup_mixed     246 μs  268 μs   +9%    148 B   161 B    +8%
lookup_most      438 μs  474 μs   +8%    308 B   293 B    -4%

Surprisingly, there are no allocations even without this change. I checked the Core and found that lookup is small enough that GHC inlines it and simplifies, removing the allocations. With this change, GHC no longer inlines lookup, which makes it slightly slower. I would argue it is still a good change because it has a positive impact on code size and compile time if lookup doesn't inline at every call-site.

[meooow25]

Point of reference: unordered-containers does the same thing, see https://github.com/haskell-unordered-containers/unordered-containers/blob/594a3055/Data/HashMap/Internal.hs#L670-L677