matmul benchmark

Benchmarking fp32 matmul kernels on 4096x4096 matrices.

• "naive" does a simple loop reduction with WebGPU block size
• "shmem-tiling" is tiled reduction with var<workgroup> memory
• "unroll4" has each thread compute a 4x4 block of output
• "unroll4x2" has 4x4 blocks of output, with 2x loop unroll
• "unroll4x4" has 4x4 blocks of output, with 4x loop unroll
• "onnx" runs a MatMul node in onnxruntime-web
• "tfjs" runs tf.matMul()
• "jax-js" runs jax.numpy.dot()