[InstCombine][X86] Try to convert BLENDV(X,Y,SHL()) -> SELECT(ICMP_SGT(0,SHL()),Y,X) (#173389)

We are cautious about converting from BLENDV intrinsics as the mask is
usually bitcast from another type, often of an entirely different width
(especially for PBLENDVB which is often used for all integer types) -
incorrect handling can leave us with select ops working on the wrong
type width, which makes it difficult for other passes to make use of it
(VectorCombine in particular).

Currently BLENDV intrinsics are only folded to generic selects when we
know the mask is from a SEXT(vXi1) bool type.

But a second common use is to shift specific bits to the MSB of the
blend mask - this is common in fp mathlib code when working with bounds
etc. and the backend is pretty good at folding this back to a
VSELECT/BLENDV pattern (often better than using the shift directly
especially when it has a non-uniform shift amount).

I've been looking for other common arithmetic ops that would benefit
from this fold, but haven't found much yet - the remainder are usually
signbit/copysign/logic manipulations that are difficult to guarantee we
have the correct element types.