X86: make VBMI2 funnel shifts use VSHLD/VSHRD for const splats

Move constant splat handling for vector funnel shifts into a DAG combiner
so that VBMI2 legal widths emit VSHLD/VSHRD directly (fixes #166949).

Signed-off-by: Arnav Mehta <[email protected]>

Author: ArnavM3434