Use gcc __builtin_bit_cast when available.

Get rid of specializations -- they are not needed, as verified by
godbolt.

Signed-off-by: Larry Gritz <[email protected]>

vectorize loops containing it:

* Restore the intrinsics, just for icc, just for the float<->int/uint
  varietes (still axe the ones involving doubles).

* Remove the __builtin_bitcast clauses, it seems to provide no benefit
  for the compilers that support it (icc doesn't, and that's the one
  that seems to need the exra hints).

* Add comments explaining why this is all the case and also reminding
  us NOT to switch to C++20 std::bit_cast in the future without also
  testing whether using it prevents auto-vectorization.

Signed-off-by: Larry Gritz <[email protected]>