[Dev] Bring Block Reduction into our seach space and policy

[LeiWang1999]

This pull request includes several changes to the bitblas module, focusing on improving functionality, optimizing performance, and refactoring code. The most important changes include adding a deprecation decorator, modifying tensorcore policies, enhancing the general_matmul module, and updating the ladder_permutate and lop3_permutate modules.

Code Enhancements and New Features:

• bitblas/__init__.py: Added a deprecation decorator to mark functions as deprecated and emit warnings when used.
• bitblas/module/__init__.py: Introduced the unpack_qweight function for unpacking quantized weights and updated the load_and_transform_weight method to use this function. [1] [2]

Tensorcore Policy Optimizations:

• bitblas/base/roller/policy/tensorcore.py: Refactored _check_small_tile logic, removed commented-out block reduction constraints, and added _expand_with_tags to handle block reduction depth. [1] [2]

General Matmul Module Enhancements:

• bitblas/ops/general_matmul/__init__.py: Added weight_compress and updated the _assign_weight_compress method to handle weight compression. Refactored transform_weight to improve weight processing. [1] [2] [3]

Ladder Permutate and LOP3 Permutate Module Updates:

• bitblas/ops/ladder_permutate/__init__.py: Added forward and retrieve_output_shape methods to improve functionality.
• bitblas/ops/lop3_permutate/__init__.py: Updated forward method to handle input and output tensors more efficiently and added retrieve_2d_weight_shape method.