Apply centered-patches optimization to the GPU path

[YQ-Wang]

What

Brings the GPU path in line with the CPU path's "centered patches" formulation and refreshes the benchmarks.

• CPU path (existing, commit db1ff4d) folds upstream's centroid term into the neighbor matrix via _centered_patches() — PatchesCells = PatchesCells_Raw - K_NN * Diagonal — so each feature batch is a single sparse matmul (PatchesCells %*% ExpMat).
• GPU path (this PR) now does the same: the CUDA sparse tensor is built from _centered_patches(...) and the per-batch diff_ik.sub_(k_nn * exp_batch_gpu) is removed.
• Removed the now-redundant call-site comments (the _centered_patches docstring already explains the rationale).
• Refreshed the README Reference performance and graph-reuse benchmarks.

Why

The change is algebraically exact:

(patches_raw - k_nn*I) @ exp  ==  patches_raw @ exp - k_nn*exp

patches_raw has a zero diagonal (self is excluded from the k-NN), so centering just sets the diagonal to -k_nn, matching upstream CastleLi/STORM exactly. Both approx and exact GPU modes share this matmul, so both are covered. The result is one unified formulation across CPU, GPU, and upstream — and one fewer per-batch temporary/kernel on GPU.

Verification (NVIDIA RTX 5070 Ti, torch 2.10 + CUDA 12.8)

• Tests: all 33 pass, including the three CPU↔GPU equivalence tests (test_storm_sparse_gpu_equivalence, test_storm_exact_gpu_equivalence, test_cached_gpu_results_match).
• Numerically identical: CPU↔GPU corr = 1.000000, max p-value/effect-size diff < 5e-5, identical p < 0.05 calls — unchanged before vs. after.
• Timing-neutral: GPU time within run-to-run noise (the removed sub_ was cheap next to the SpMM + float64 column reductions).
• Memory-neutral: peak GPU VRAM changes by sub-MB, scaling linearly with cell count (the one extra -k_nn diagonal nonzero per row); the removed per-batch temporary wasn't the high-water mark (the cuSPARSE SpMM is).

Reviewer notes

• The PR also includes the prior CPU-centering commit db1ff4d already on this branch.
• README benchmark numbers were re-measured (median of five) and the graph-reuse paragraph updated accordingly.