Long vectors might not performant efficient as vectors of 4 elements

[lpavank]

This issue stemmed from the discussion here https://github.com/google/clspv/issues/1451#issuecomment-2677502673 with @rjodinchr . Given the fact that, 4 element vectors are natively supported by Vulkan SPIR-V vs this is not the case for 8 or 16 element vectors. Functionally long vectors (i.e. vectors of 8, 16) are ok, but performance might suffer. This task/issue to keep track if any can be done in cases where kernels can't be changed (4 element vectors for perf)

PS: Yes, I don't have data at this time to quantify how bad it is at this time, but the discussion mentioned above is a data point.