Add reduce_sum op

[Tcc0403]

This PR adds a simple reduce_sum kernel, which can be a starting point for new starters.

Strategies

• leveraging TV layout for vectorized load
• utilizing warp reduction intrinsic to avoid syncrhonization and reduce memory pressure
• skipping shared memory if reduction dimension is small
• performing reduction in higher precision for numerical stability

Benchmark results

I only did a simple benchmark on my local machine, and it already performs better than torch's.

On RTX 3080

❯ uv run bench/benchmark_reduce_sum.py --m 4096 --n 4096 --compile-ref
copy_transpose p50: 0.0655 ms, BW: 512.13 GB/s
reference p50: 0.1024 ms, BW: 327.76 GB/s

Limitations

• hardcoding num_threads and threads_per_row, which is not scalable across different input shapes
  • heuristic approach: https://github.com/Dao-AILab/quack/blob/9e774f8225fa890aa0654a9a2350991407b9f2c2/quack/softmax.py#L34
• only support dim=0
• assumption on input layout. Input tensor has to be contiguous
  • To address above 2 points, it requires different tv_layout for colalesced, vectorized memory access
• ouputs might be incorrect with small input shapes (only tested with (M, N) where M is mutliple of 4 and N is 1024)
  • needs predication, see Predication: What to do when tiling isn’t perfect

Future works

• adding all common reduction ops that torch has, e.g. torch.amax, torch.amin, torch.prod, torch.mean
• eliminating all limitations

related issue: #20