[vectorapi] Optimize masked store for non-predicated architectures

Currently the vectorization of masked vector store is implemented by the masked store instruction only on architectures that support the predicate feature. For non-predicate supported architectures, the masked store can also be vectorized with "load + blend + store".
For example, implementation for "store(v, m)" can be implemented with:
1) v1 = load()
2) v2 = blend(v1, v, m)
3) store(v2)