nir/opt_vectorize: small improvements for speed and around swizzles

This series adds a couple of patches to avoid hashing of instructions which cannot be vectorized. Additionally, swizzled accesses to components higher than the maximum vectorization width are also hashed and compared. The idea is that vectorized instructions should only access components within the number of components it can process.

The array of instructions per hash entry is removed. For this change, I'd like to get some feedback if it has any negative impact on the results.