[Vector] Vector distribution (large vector to small vector)

[nicolasvasilache]

Slowly unpacking things post-vacation…
This seems mostly resolved given the phab review I saw, I’ll just reply to the transformation part.

@mehdi_amini

If the intent is that these operations cannot exist in the IR in between passes, I would rather avoid having these as part of the vector dialect spec and documented “as if” they have a well defined semantics.
They would be really just private markers for the internal state of a particular transformation. (In the extreme we could even have their verifier to unconditionally fail, but that’d be a drag on testing/development though).

They should be able exist in the IR, similarly to other ops. For envisioned use cases, we want to continue investing in progressive lowering and pattern composition with more dynamic semantics on vectors. But we may also want at some point to have them lower directly to LLVM (even if very inefficiently). I agree though that if the op could never materialize in the IR then attributes or C++ would be a better encoding.