Add all experiment code

triton_sparse.py

@@ -0,0 +1,15 @@
+#!/usr/bin/env python3
+"""
+Triton-fused Chunked Sparse Backward Pass.
+
+Replaces the Python for-loop over active chunks with fused Triton kernels:
+  1. sparse_bwd_dW: grad_W[c*CS:(c+1)*CS, :] = grad_Y[:, c*CS:(c+1)*CS].T @ X  for active c
+  2. sparse_bwd_dX: grad_X += grad_Y[:, c*CS:(c+1)*CS] @ W[c*CS:(c+1)*CS, :]  for active c
+  3. sparse_bwd_dbias: bias_grad[c*CS:(c+1)*CS] = dY[:, c*CS:(c+1)*CS].sum(dim=0)
+
+Includes Python-loop baseline, correctness tests, and isolated matmul microbenchmark.
+
+Usage:
+    python triton_sparse.py   # runs correctness + benchmark
+"""
+# See repo file for full content - uploading from sandbox