Reduce sum implementation

[vvvdwbvvv]

Implement CuTe Reduction Kernel with Multi-Variant Support

Related to #20

---

What Changed

This PR implements a high-performance row/column sum reduction kernel using NVIDIA CuTe DSL, with multiple optimization variants based on Mark Harris's parallel reduction techniques.
Optimizations (based on Mark Harris reduction techniques):

1. Algorithm Cascading: Each thread processes multiple elements sequentially
   before reduction, reducing instruction overhead and enabling better latency hiding.
2. Warp Shuffle: Use warp_reduction_sum to avoid shared memory for intra-warp reduction.
3. Two-layer reduction: thread-local → warp shuffle → inter-warp via shared memory.
4. Vectorized loads: Use 128-bit loads for memory bandwidth efficiency (row reduction).

Features Added

Feature	Description
Bidirectional reduction	Supports both dim=0 (column) and dim=1/-1 (row)
Multi-variant	naive, improved, shfl optimization levels
Dtype support	float16, bfloat16, float32
Kernel caching	Compiled kernels cached by (dtype, dim, variant, M, N)

Optimization Variants

Variant	Vecsize	Threads/Row	Description
naive	1	32 (1 warp)	Baseline, no vectorization
improved	128-bit	32 (1 warp)	Vectorized loads
shfl	128-bit	up to 128 (4 warps)	Multi-warp + warp shuffle

Limitation

• Support 2D
• Currently only support sum
• Input has to be contiguous
• Dim=0 is implemented via transpose

---

Test Done

Environment

OS:  Ubuntu 24.04 LTS
GPU: NVIDIA RTX 3090
CUDA: 13.0
Python: 3.13.11

Test Commands

# Run all reduce_sum tests (91 tests)
uv run pytest tests/test_reduce_sum.py -v

# Run pre-commit checks
uv run pre-commit run -a

Test Coverage

Test output

=================== 91 passed, 14 warnings in 17.26s==========================

---

Bench Results

Benchmark commands

# Run benchmark
uv run bench/benchmark_reduce_sum.py --m 4096 --n 4096 --dtype float16

# All dtypes
for dt in float16 bfloat16 float32; do
  echo "=== dtype=$dt ==="
  uv run bench/benchmark_reduce_sum.py --m 4096 --n 4096 --dtype $dt
done

---

Benchmark results

=== dtype=float16 ===
copy_transpose p50: 0.0501 ms, BW: 670.39 GB/s
reference p50: 0.2222 ms, BW: 151.04 GB/s
=== dtype=bfloat16 ===
copy_transpose p50: 0.0486 ms, BW: 690.02 GB/s
reference p50: 0.2222 ms, BW: 151.04 GB/s
=== dtype=float32 ===
copy_transpose p50: 0.0881 ms, BW: 762.23 GB/s
reference p50: 0.0932 ms, BW: 720.35 GB/s