Lift scalar VABS directly when the lane and register widths match, and
use intrinsics for packed ARM and Thumb forms

Correct Thumb F32/F64 VCVT result sizing, fix scalar fixed-point
rounding and extension, and add an intrinsic for lane-wise fixed-point
conversions