feat(csrc): Optimized inference kernel for vector-quantized gemv.

[lcy-seso]

This pull request is a work in progress aimed at creating an optimized inference kernel for the vector-quantized GEMV operation.

The operation follows these steps:

1. Load the codebook into shared memory.
2. Load tiles for inputs, indices, scales, and bias from shared memory into registers.
3. Decode the packed indices and dequantize the weights using lookup operations.
4. Compute the GEMV by performing a reduction sum using warp-level primitives.
5. With CUTLASS now added as a dependency, update setup.py to automatically check it out as a submodule.

Toward the final goal, this pull request still requires additional unit tests to ensure the correctness of the results. These tests will be added in subsequent pull requests.