[AArch64] Pre-commit tests for #125686 (NFC)

Update the `generate-tests.py` script to create new tests for
`atomicrmw {fadd,fmin,fmax}` and test these with `half`, `float`,
`bfloat` and `double`.

Generate auto-tests to check both with and without `+lsfe`, so
that when #125868 is merged, `+lsfe` will use a single atomic
floating-point instruction.

Author: jthackray