[LLVMGPUVectorDistribute] VectorDistribution support for unaligned shapes

[Groverkss]

This PR adds support to perform statically tiled codegen on dynamic shapes in vector distribute pipeline.
Basically, it could honor lowering configs on dynamic shapes using masking.

Some side-effect changes:

• Currently block dynamic dimension pass, change the dimensionality of the generics without performing a projection of the lowering config that was provided higher up in the pipeline. Moreover, the requirement to do this becomes less now as we can tile generally on the dynamic dimension with the changes here -- unless Im missing something here.

This builds on the following PRs -- hence putting to draft:

• [LLVMGPUVectorDistribute] Add support for distributing masked reads/writes #19830
• [LLVMGPUVectorDistribute] Support vector.mask + vector.multi_reduce #19880
• [LLVMGPUVectorDistribute] Support vector.mask + vector.contract  #19899
• [LLVMGPUVectorDistribute] Infer sizes for dynamic shapes from lowering config #19973

future work:

• (near future) Enable intrinsic distribution with selects.
• (near future) Once [MLIR][Vector] Add support for inner-parallel masked multi-reductions llvm/llvm-project#126722 is merged and integrated, we can enable masking post-distribution.
• [AMDGPU] once buffer descriptor based pointers are properly introduced, we could rewrite the masked loads to be just loads followed by selects. (This works with the prototype : [Codegen][ROCDL] Use buffer fat pointers where possible #19918)

Original Author: @manupak

[AmosLewis]

Just test llama3 benchmark on input size length 128/2048 by locally rebased this PR on 3.3.0rc20250310. Improve performance around 1ms.
128: 37.2 ms->36.4 ms.
2048: 174 ms -> 173 ms.

[Groverkss]

Just test llama3 benchmark on input size length 128/2048 by locally rebased this PR on 3.3.0rc20250310. Improve performance around 1ms. 128: 37.2 ms->36.4 ms. 2048: 174 ms -> 173 ms.

This should have no effect on any benchmarks...