fix(cuda): bound DeepSeek V4 prefill memory

[AlpinDale]

Summary

• write SM89 sparse-prefill output directly into the engine buffer
• reserve one shared DeepSeek V4 padded-query workspace per worker
• remove request-time padded-query allocation from the fused CUDA kernel

Validation

• touched-file pre-commit checks pass
• prior output-buffer fix survived 9 uncached 70k-token requests in 3 concurrent batches on 8x RTX 4090
• PR wheel will be used to repeat the stress test with both allocation fixes

[AlpinDale]

/build-wheels