[X86][SSE] Prefer to combine shuffles to VZEXT over VZEXT_MOVL.

This matches what is already done during shuffle lowering and helps prevent the need for a zero-vector in cases where shuffles match both patterns.

llvm-svn: 295723