feat(deepseek-v3.2 / glm-moe-dsa): implement the DSA lightning indexer

[inureyes]

Summary

Replaces the full-attention fallback in deepseek_v32 (DeepSeek V3.2) and glm_moe_dsa with the DeepSeek Sparse Attention (DSA) "lightning indexer" these architectures were trained for, so the models attend to the top-k most relevant KV entries per query and restore long-context fidelity.

Changes

• Shared indexer module (src/models/deepseek_v32_indexer.rs): the wq_b / wk / k_norm / weights_proj projections, relu(q @ k^T) scoring combined with the per-head weights_proj scale, causal masking, and top-index_topk key selection via gather.
• Per-model indexer_rope_interleave flag threaded into the indexer RoPE, with the correct per-model default: non-interleaved (false) for deepseek_v32, interleaved (true) for glm_moe_dsa. Getting this wrong silently corrupts key selection, so it is covered by explicit tests.
• Config plumbing: index_topk / index_n_heads / index_head_dim / indexer_rope_interleave added to deepseek_v32 (glm already parsed index_*), and the wq_b / wk / k_norm / weights_proj indexer weights are loaded. glm_moe_dsa's previously-unused index_* fields are now wired.
• When kv_len <= index_topk the indexer reduces to the dense path (short context stays numerically equivalent to the previous fallback), and the L==1 decode path gathers the selected keys instead of materializing a per-step sparse mask.

Correctness (covered by unit tests)

• deepseek_v32_indexer_rope_defaults_to_non_interleaved and glm_moe_dsa_indexer_rope_is_interleaved_by_default: the per-model RoPE interleave default is correct.
• indexer_load_threads_rope_flag_per_model, glm_moe_dsa_falls_back_to_direct_rope_fields, glm_moe_dsa_prefers_rope_parameters_over_direct_fields, ..._config_overrides_are_respected, ..._config_defaults_thread_through.
• indexer_load_absent_weights_is_dense_fallback and indexer_top_indices_none_at_short_context: reduces to dense when weights are absent or when kv_len <= index_topk.
• indexer_top_indices_selects_highest_scoring_keys and indexer_top_indices_respects_causal_mask: top-k selection is correct and causal.

Validation

• cargo check --lib --tests --features metal,accelerate: compiles clean.
• cargo test --release --features metal,accelerate deepseek_v32: 7 pass. ... glm_moe_dsa: 5 pass. Total 12 indexer unit tests pass on Apple Silicon Metal.
• Not run here: generation on a real DeepSeek-V3.2 or GLM-MoE-DSA checkpoint at context length greater than index_topk. Those real-model parity tests are gated ignored (they require local model weights and, for the pipeline variants, TCP-bound remote stages), and no such checkpoint is present in this environment. Long-context behavior is validated at the indexer scoring / top-k selection / dense-reduction level, not real-model generation, and should be confirmed on a real checkpoint before relying on it in production.

Closes #509