Feat/profiling

[nathanielsimard]

Two level of profiling for now:

Basic

cmd: CUBECL_DEBUG_OPTION=profile CUBECL_DEBUG_LOG=stdout cargo bench --bench matmul --features cuda-jit
For 5 kernels:

| 7.438437ms | CmmaKernel
| 6.712298ms | CmmaKernel
| 6.683767ms | CmmaKernel
| 7.825637ms | CmmaKernel
| 7.125788ms | CmmaKernel

Full

cmd: CUBECL_DEBUG_OPTION=profile-full CUBECL_DEBUG_LOG=stdout cargo bench --bench matmul --features cuda-jit

For 1 kernel:

| 7.079846ms | cubecl_linalg::matmul::cmma::base::cmma_kernel::CmmaKernel<half::binary16::f16, half::binary16::f16, cubecl_cuda::runtime::CudaRuntime>: (
    KernelSettings {
        mappings: [],
        vectorization_global: None,
        vectorization_partial: [
            Input {
                pos: 0,
                vectorization: Some (
                    8,
                ),
            },
            Input {
                pos: 1,
                vectorization: Some (
                    8,
                ),
            },
            Output {
                pos: 0,
                vectorization: Some (
                    8,
                ),
            },
        ],
        cube_dim: CubeDim {
            x: 32,
            y: 8,
            z: 1,
        },
        reading_strategy: [],
    },
    ComptimeCmmaInfo {
        block_size_m: 128,
        block_size_k: 16,
        block_size_n: 128,
        tile_size: 16,
        check_m_bounds: false,
        check_k_bounds: false,
        check_n_bounds: false,
        unroll: false,
        coop_dim: 32,
        num_coops: 8,
        num_accumulators: 8,
        write_out_strategy: 1,
        cube_dispatch_strategy: 1,
        compute_loop_order_strategy: 0,
        reuse_lhs_fragment: false,
    },
) CubeCount (16, 16, 8)