Add indexCache training support

[faresobeid]

Can use with

[trainer.model]
impl = "custom"
use_index_cache = true
index_topk_freq = 4

[inference.vllm_extra.hf_overrides]
use_index_cache = true
index_topk_freq = 4

---

Note

Medium Risk
Touches core attention forward paths and introduces cross-layer state (cached indices), which could affect correctness/performance if misconfigured or if assumptions about layer scheduling break.

Overview
Enables DSA IndexCache in training by adding use_index_cache, index_topk_freq, and optional index_topk_pattern to the trainer ModelConfig and propagating them into the loaded HF model_config.

Updates the custom glm_moe_dsa implementation so decoder layers can reuse sparse attention top-k indices across layers: attention/decoder forwards now thread a cached_indices tensor through the stack, and a new per-layer skip policy (_index_cache_skip_topk) controls when indices are recomputed vs reused.

^{Reviewed by Cursor Bugbot for commit 563b77f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 563b77f. Configure here.}