fix: restore Kimi Linear CUDA inference

[inureyes]

Problem

Kimi Linear treated the repository's additive causal mask (0 for visible keys and negative infinity for blocked keys) as a boolean condition. This reversed key visibility. In every multi-token prefill, the final query row contains only zeros, so the implementation replaced all of its positional scores with negative infinity. Softmax then produced NaNs, which propagated through later hybrid layers and made greedy decoding emit token ID 0 repeatedly.

Change

• Cast the causal mask to the MLA positional-score dtype and add it to the scores, matching the mask contract and the Kimi K3/DeepSeek MLA implementations.
• Add an exact mask-value regression and a loaded three-token MLA prefill regression that requires finite output.
• Record before/after results for the pinned Kimi Linear checkpoint in bilingual technical reports.

Validation

• cargo fmt --all -- --check

• git diff --check

• Root and surgery CUDA test-fast build graph: passed.

• Focused final regressions: 3 passed, including the two MLA mask tests and Kimi's loaded-model chunk capability test. Replacing only the mask helper with the former where_cond semantics made both MLA tests fail as expected.

• Shared issue contracts: 123 Python checks and 49 WebUI checks passed with exit 0.

• Pinned checkpoint mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit at 919175d75427f1c10cfc17e7254b53295edddf72: the 21-token prompt returned exactly 4; the 3,411-token prompt returned exactly AMBER-RAVEN-7429 with MLXCEL_PREFILL_CHUNK unset and set to 0. The ordinary MLXCEL_SDPA_DETERMINISTIC=0 short and long paths returned the same exact answers.

• Additional ordinary-path generation (MLXCEL_SDPA_DETERMINISTIC=0, prefill override unset) produced a coherent 58-token, two-sentence explanation of daytime blue sky and reddish sunsets; this qualitative smoke test used the same pinned checkpoint and tested binary.

Closes #2292