Implement BlockPointerToTensorDesc pass

[whitneywhtsang]

This PR implements the BlockPointerToTensorDesc pass, which translates block pointer representations to tensor descriptor pointer representations. As part of the planned deprecation of block pointer representations, this pass serves as a transitional solution. Since many existing kernels still use block pointers, we cannot immediately remove block pointer support from the compiler. To simplify the compiler architecture and avoid maintaining dual pointer representation handling, this pass converts block pointers to tensor descriptors, allowing us to standardize on a single pointer representation throughout the compilation pipeline.

Note: Currently, the pipeline immediately translates tensor descriptors back to block pointers. However, we plan to incrementally update each compilation step to handle tensor descriptors directly, eventually eliminating the need for block pointer representations entirely.

Fixes #5784.

No geomean performance regressions:
Triton Benchmarks PVC: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20633795972
Triton Benchmarks BMG: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20633799845 + https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20642028915

[etiotto]

Please link the issue to the PR.

[whitneywhtsang]

Please link the issue to the PR.

done

[etiotto]

Initial comments.

[whitneywhtsang]

To check:
PyTorch core: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20833194817
PyTorch FlexAttention: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20833147887
PyTorch FlexDecoding: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/20833167619