[AArch64] Improve fsh(l|r) cost modeling if 3rd arg is constant.

In that case, the cost for i32 and i64 should be 1 (a single EXTR
instruction). For v4i32 and v2i64 it should be 3 (USHR + SHL + ORR).

Other sizes smaller than 64 bits require an extra instruction for
conversion to i32/i64.

This recovers a SLP regression revealed by D140392.

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D147322