Attention optimizations (part 2): channel-aware backward cache

[azrael417]

Summary

Second batch of spherical-attention work on top of the optimized kernels. User-visible piece:

• Channel-aware backward cache for the serial neighborhood-attention CUDA backward. The two-pass flash-style backward previously re-gathered K/V and recomputed qdotk/gdotv in pass 2. For wide
  channel counts we now cache those per-neighbor dot products in shared memory in pass 1 and read them back in pass 2, eliminating the recompute. Gated to nchan >= 192 and compile-time templated
  (template + if constexpr) so the sub-192 path is byte-for-byte identical to the previous kernel — no effect on the common C=64 ERA5 workload.

Plus benchmark and test coverage for the wide-channel regime the cache targets.

Motivation

The backward pass was the remaining hotspot after the forward work in part 1. Profiling showed pass-2 recompute (re-gathering K/V and redoing the dot products) is pure redundant work when
there's shared-memory budget to hold the pass-1 results — which only pays off once the channel dimension is wide enough to dominate the per-neighbor cost, hence the nchan >= 192 gate.
WGMMA/tensor-core rewrites were prototyped and rejected (small skinny latency-bou Jump to bottom (ctrl+End) ↓ r), so this MR keeps the scalar/vectorized kernel and only removes redundant

This MR is marked as draft because #231 needs to be merged first.