Adding a memref linearizer pass

[lialan]

This PR responds to #20226

This PR is based on the GPU/DecomposeMemRefs pass in upstream, with the following changes:

• support vector dialect ops
• instead of decompose memrefs to rank-0 memrefs, changed to decompose memrefs to rank-1.

Notes:

• when calculating linearized offsets, it dissolves memref's offset and make it part of the linear index calculation. So it is possible to result in out-of-bound accesses.

What is next?

• Refactor and push back to upstream.
• consider move the upstream pass from GPU to common.
• use it as prerequisite for passes such as EmulateNarrowTypes.

[krzysz00]

High level comment: we might want - like we have in narrow type emulation - explicit rewrites for alloc() and friends (and so, IREE-side, hal.interface.buffer_subspan)