cuda: accelerate long-context B1 indexer and 1M HCA

[wangshang23]

Fixes #762.

What changed

• Route non-quality B1 indexer scoring to the existing WMMA scorer at
  n_comp >= 8192; retain the direct scorer for shorter contexts, quality
  mode, and the existing DS4_CUDA_NO_INDEXER_WMMA escape hatch.
• Raise the attention score capacity from 8192 to 8448 rows so the 1M HCA
  shape fits 8192 compressed rows plus the existing 256-row raw reserve.
• Extend the CUDA long-context smoke test to cover 8192 compressed HCA rows
  and direct-vs-WMMA B1 score/top-k equivalence on QAT-format values.

The HCA change does not add a new attention implementation. It keeps the
existing exact score-split path active instead of falling back to the slow
online kernel for the last 256 rows at 1M.

DGX Spark results

Base: 84cc882, NVIDIA GB10 (sm_121), DeepSeek-V4-Flash IQ2XXS Q2 model,
Q8-to-F16 weight cache disabled. A/B paths alternate in one process/session,
with 3 warmups and 30 samples each.

Context	Base tok/s	PR tok/s	Speedup
128K	12.168	12.468	1.025x
256K	10.642	11.290	1.061x
512K	8.647	9.574	1.107x
1M	4.372	6.981	1.597x

At 1M, Nsight Systems attributes the main change to:

• CSA: direct scorer 78.036 ms -> WMMA scorer 48.565 ms across 21 layers.
• HCA: online kernel 97.286 ms -> exact score-split score+finalize
  25.400 ms across 20 layers.

Full benchmark details and the existing-work audit are in #762. Related #258
uses approximate HISA pruning and is complementary to this full-top-512 flat
scorer optimization.

Correctness

• B1 QAT score max delta: 6.10352e-05.
• B1 top-512 differences: 0/512.
• Isolated 8192-row HCA regression output delta: 0.
• Full synthetic-frontier logits are bit-identical through 512K.
• At 1M, legacy online vs existing exact score-split has max logit delta
  0.0469046, RMSE 0.00876944, and identical argmax.
• Real 30,474-token story fact-recall regression passes.

Validation

make cuda-regression CUDA_ARCH=sm_121

Passes on DGX Spark / GB10.

--logprob-vectors currently fails short_code_completion identically on
both base and PR builds with this GGUF, so it is not a regression from this
change.