Fix Hash#reject! for non-equality key

[straight-shoota]

Fixes the immediate issue reported in #3923. It does not decide the discussion on hashing of NaN values but ensures equal behaviour between #reject and #reject! (and by extension, #select and #select!).

Currently, you get this behaviour because #reject! deletes by key which is based on quality of the key value:

h = {Float64::NAN => true}
h.reject(&.nan?)  # => {}
h.reject!(&.nan?) # {Float64::NAN => true}

With this change the latter call also properly deletes the entry.

I would expect this change to also improve performance because it avoids a duplicate hash lookup on deletion, but I couldn't actually demonstrate that in a benchmark. Performance is actually stable.
Maybe I'm missing something.
But it fixes a bug and doesn't deteriorate performance, so it's fine to merge.