Issue 20 reduce sum

[debashishc]

PR up: reduce_sum kernel + benchmark cleanup. Better late than never, cleaned up the implementation a bit. Working on reducing for loops next.

Highlights:

• Vectorized 128‑bit loads (CopyAtom + TV layout)
• GMEM→SMEM via cp.async, then SMEM→registers
• Thread‑local reduction + warp shuffle reduction
• Block reduction via shared memory
• Alignment/divisibility constraints for 128‑bit copies
• Removed explicit FP32 accumulation in ref to cut overhead

Next step: remove the tail loop via a masked copy path to keep vectorized loads across the full row.

Refs #20