[llvm] [LLVM] Add `llvm.masked.compress` intrinsic (PR #92289)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

If FoldConstantArithmetic is already special casing ISD::CONCAT_VECTORS I don't see why we shouldn't just special case another generic ISD as well - we could create a ISD wrapper if you prefer e.g. ISD::isVectorOnlyOp(unsigned Opcode) ?

https://github.com/llvm/llvm-project/pull/92289