Avoid maximum in softmax

[mcabbott]

I think max is very careful about NaN & signed zero, and hence maximum is surprisingly slow. This uses the @fastmath version:

julia> let x = randn(Float32, 100, 1000)
         @btime softmax($x)
         @btime _softmax($x)  # this PR
       end;
  min 1.344 ms, mean 1.422 ms (7 allocations, 398.91 KiB)
  min 647.875 μs, mean 707.579 μs (3 allocations, 394.73 KiB)

[CarloLucibello]

We could remove the doctest

[Sleort]

The new fast_maximum isn't always faster than maximum, ref. #456 (comment) ...