[Feature] Kernel for sparse matrix-dense matrix multiplication

[mercush]

Proposed changes

Feature: added function metal kernel sparse_matmul_csr for sparse matrix-dense matrix multiplication. The function takes as input a sparse matrix (represented as three arrays representing the row pointers, the column indices, and the values at the nonzero entries) and a dense matrix represented as a dense array with two dimensions. I have implemented the backends for metal and for the cpu but not cuda.

Please let me know if there is any documentation I should update.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[mercush]

Metal, Cuda, and cpu backends are passing tests.

[mercush]

Benchmark: spmm can be a lot faster for large matrices with a lot of sparsity.

Metal backend (M2 Pro):

Sparse matmul CSR benchmark (1024x1024, 5% nonzero)
  float16:
Timing sparse (float16) ... 0.42803 msec
Timing dense  (float16) ... 0.65754 msec
  bfloat16:
Timing sparse (bfloat16) ... 0.38357 msec
Timing dense  (bfloat16) ... 0.64950 msec
  float32:
Timing sparse (float32) ... 0.49228 msec
Timing dense  (float32) ... 0.60805 msec
Sparse matmul CSR benchmark (4096x4096, 1% nonzero)
  float16:
Timing sparse (float16) ... 6.89379 msec
Timing dense  (float16) ... 22.62993 msec
  bfloat16:
Timing sparse (bfloat16) ... 6.82182 msec
Timing dense  (bfloat16) ... 26.86476 msec
  float32:
Timing sparse (float32) ... 15.21455 msec
Timing dense  (float32) ... 27.31441 msec

CUDA backend (RTX 5090):

Sparse matmul CSR benchmark (1024x1024, 5% nonzero)
  float16:
Timing sparse (float16) ... 0.05121 msec
Timing dense  (float16) ... 0.05960 msec
  bfloat16:
Timing sparse (bfloat16) ... 0.04969 msec
Timing dense  (bfloat16) ... 0.06564 msec
  float32:
Timing sparse (float32) ... 0.06870 msec
Timing dense  (float32) ... 0.07317 msec
Sparse matmul CSR benchmark (4096x4096, 1% nonzero)
  float16:
Timing sparse (float16) ... 0.24195 msec
Timing dense  (float16) ... 0.72532 msec
  bfloat16:
Timing sparse (bfloat16) ... 0.23913 msec
Timing dense  (bfloat16) ... 0.84904 msec
  float32:
Timing sparse (float32) ... 0.44731 msec
Timing dense  (float32) ... 1.44518 msec