qwen4exp: fix per-block bias indexing when a unified cache holds several sequences

[akionux]

Overview

Fix the per-block (blk_bias) path of set_input_qsa: it indexes the per-sequence bid arrays by block number (bid_cell[b], bid_idx[b]), but those arrays are indexed by bid, and the block table keys a block on (sequence set, bucket). With more than one sequence in a unified cache the two no longer coincide, so a block gets the visibility of a different sequence's group - one of our blocks can be masked out, one of another sequence's can pass through.

Build an inverse block -> bid map once per ubatch, keeping the entry that owns the stream's sequence, and use it for both the seq_has test and the tail comparison. Single-sequence caches are unaffected (one_seq keeps the old correspondence).

Reproduced with two live sequences in one stream (llama-server serves each slot as its own sequence from position 0): the model stops seeing its last messages and reports the input as empty. With the change, the same concurrent shape is answered correctly in 4/4 attempts (2/4 before).

Additional information

Found while running this branch's QSA path under --parallel 4 with concurrent requests. The companion fixes for the pooled-key-cache-specific paths (per-sequence row windows, own-block fill, stream-inventory sizing) live in #28699; this PR is the part that exists on master on its own.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - patch drafted by an AI agent (Hermes) under the author's direction on the author's account; the author is responsible for all submitted changes, per AGENTS.md/CONTRIBUTING.md.

[ggml-gh-bot]

Hi @akionux, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.