chore: bump mlx-swift-lm to include sortedIndices fix for stacked MoE path (#79)

[solderzzc]

Summary

Bump the mlx-swift-lm submodule to current fork main. No SwiftLM source change, and the mlx-swift pointer is unchanged (3764829).

• mlx-swift-lm a8c85d5 -> 9da07be (fix(moe): stacked LRU slot ids must not be declared sorted (mlx 0.32) mlx-swift-lm#79): the opt-in MLX_MOE_STACKED fused path no longer declares non-ascending LRU slot ids as sortedIndices. mlx core 0.32 requires ascending ids for sortedIndices: true, and with unsorted ids the result is wrong at n >= 64 rows. This is defensive: the stacked path is gated to idx.size <= 32 and is opt-in, so default runs were not affected. feat: bump mlx-swift-lm for DeepSeek-V4 support #79 includes a regression test.

Test plan

• Checked against the real Qwen3.5-35B-A3B-4bit with --stream-experts: no regression.
• Not exercised: the stacked path itself (MLX_MOE_STACKED=1) was not run on the real model. Its coverage is the regression test added in feat: bump mlx-swift-lm for DeepSeek-V4 support #79.

CI:

• build_and_unit_test
• Remaining integration_matrix jobs

AI disclosure

This PR was prepared with AI assistance (Claude Code). It still needs human review before merging.

• I have read this PR description in full and approve it as my own

🤖 Generated with Claude Code