[AArch64][GlobalISel] Remove fallbacks for fpcvt intrinsics with 16-bit operands (#179693)
Previously, GlobalISel failed to lower neon fpcvt intrinsics, as
RegBankSelect was not keeping the result on a fpr.
An additional fix is needed for the fpcvtz intrinsics, as these are the
"default" floating point convert intrinsics. As a result, Instruction
Selection has patterns mapping the FPCVTZ intrinsic to the
architecture-agnostic G_FP_TO*I_SAT node.
This also provides the opportunity for more optimisations to be made to
the code before Selection.