[deepseek_v4] fix CSA per-query masking in eager path
Two related bugs in DeepseekV4Indexer / DeepseekV4CSACompressor that
made multi-token forwards (S > 1) silently use the wrong CSA attention
pattern:
* Indexer.topk had no causal pre-mask, so query t could select compressed
blocks containing tokens at position >= t. Add the same -inf mask +
-1 sentinel that DeepSeek's reference inference/model.py uses.
* CSACompressor flattens gathered top-k entries to [B, 1, S*k, D] and
the caller right-padded the attention mask with 0.0, so every query
attended to every other query's selection. Build a per-query block
bias (-inf off-diagonal, with the sentinel routed through) and use it
instead of the zero pad. HCA compressor return signature aligned to
(kv, None) for parity.
