Floating-point abs on AArch64 NCG

Currently AArch64 NCG generates "branchful" assembly for floating-point abs. Consequences are

• potentially slow code
• it does not handle the sign bit of NaN.

Ideally, the NCG should just emit FABS instruction.

Relevant code:

• Handling of FloatFabsOp/DoubleFabsOp
• genericFabsOp
• By the way, there's a typo in handling of MO_F32_Fabs (this is currently an unreachable code)