Like #36188, but now for extrema.

This patch also switches the implementation of extrema to use mapreduce. It improves the performance even for simple arrays:

julia> @btime extrema($(randn(1000)));
  5.502 μs (0 allocations: 0 bytes)      # before (1.6.0-DEV.220)
  2.988 μs (0 allocations: 0 bytes)      # after (7b0e72efc8f7fbf2c0bd5d30a8b6e2e22288547f)

The improvement is much more drastic when mixing it with iterator transforms because it hits the transducer path.

julia> @btime extrema($(Iterators.flatten(((0, 1, 2), randn(1000)))));
  195.393 μs (5489 allocations: 133.03 KiB)  # before (1.6.0-DEV.220)
  3.469 μs (0 allocations: 0 bytes)          # after (7b0e72efc8f7fbf2c0bd5d30a8b6e2e22288547f)

Of course, for arrays, we might want to use the same unrolling trick as in minimum/maximum. But that's tracked by #31442 and can be done later. I think this PR would be a good generic fallback that is useful as a baseline for #31442.

One thing not so great about this PR is that I had to copy the old iterate-only extrema implementation to compiler.jl as apparently it is required for bootstrap. I think it might be nice to separate out a very minimal portion of reduce.jl so that it can be included from the compiler. But that's a bit too large refactoring to do in this PR.