[AArch64] Pre-commit tests for llvm#125686 (NFC) (llvm#126643)

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 fp auto-tests to check both with and without `+lsfe`, so that when
llvm#125686 is merged, `+lsfe` will use a single atomic floating-point
instruction.

Author: jthackray