[ARM] Split vector store of constantpool loads (#242865)
Given a store(buildvector(constants)) it can be more efficient to
materialize the constants and store individually with stm / strd, to
avoid the extra constant pool load. The SLP vectorize will generate such
a pattern and instead of adjusting the cost, this adds a combine to
scalar to address all examples where this can be generated. Loop
invariant loads can be excluded, and whilst the cost model isn't perfect
but can be improved in the future if needed.