Split reduction dim pass

[adam-smnk]

Adds a pass to allow tiling contraction's innermost reduction dimension using serial loop with in-place accumulation.

Compared to other available transformations, this rewrite computes reduction sequentially with in-place accumulation which avoids temporary allocation and separate reduction operation. This tiling strategy is more friendly in terms of register and memory pressure more suitable for low-level GPU kernel generation. Similarly, restriction to the innermost dimension is there to simplify both usage and pass logic as the rewrite is geared toward progressive GEMM lowering.

Additionally, a GPU vectorization control flag is added to allow grouping of passes which target lowering through vector operations and might not be compatible with other existing lowering strategies.
Effectively, this pass perform separate K-dim split which is currently baked into linalg-to-xegpu lowering, and the two are not fully compatible.

[adam-smnk]

Now that I think more about it, it should be possible to largely simplify this logic using scf::tileUsingSCF.
The concept of the pass can probably stay the same but the whole loop creation could be delegated to TilingInterface.

[adam-smnk]

Now that I think more about it, it should be possible to largely simplify this logic using scf::tileUsingSCF. The concept of the pass can probably stay the same but the whole loop creation could be delegated to TilingInterface.

Yup, that was really straight forward. Overall the pass behavior remains the same but now named ops are preserved after tiling.

[shahidact]

Thanks, looks good to me.