feat(gguf): surface MLA metadata for DeepSeek-V2/V3 + Kimi K2 — closes #67

[mvkorobkov]

Summary

Closes #67. With the MLA absorption work in #96 already merged, the last missing piece for Kimi K2 (and any DeepSeek-V2/V3) extraction from GGUF was reading the MLA geometry off the GGUF metadata. `to_config_json` dropped every `attention.q_lora_rank` / `attention.kv_lora_rank` / `attention.key_length[_mla]` / `attention.value_length[_mla]` / `rope.dimension_count` key it saw, so `ModelConfig.qk_nope_head_dim` etc. came back as `None` and `uses_mla()` stayed `false` for GGUF-sourced models.

Surprise from looking at real Kimi K2.6 files

Inspecting Kimi-K2.6 UD-Q8_K_XL (unsloth's 554 GB 14-shard split) with `gguf-dump` showed that the "Q8_K_XL" naming is misleading — the tensor type histogram across multiple shards is only BF16 + F32 + Q4_0, every one of which is already covered by larql's existing ggml dequant. No Q8_K (type 15) dequant work was actually needed for this family. The blocker was purely the missing config plumbing.

(For posterity: shard 2 = 54 BF16 / 33 F32 / 13 Q4_0; shards 3, 7, 10 = same pattern with 45/30/15; shard 14 trails with 2/2/2 — all supported.)

What this PR changes

In `crates/larql-models/src/loading/gguf.rs::to_config_json`:

GGUF key	HF field surfaced	Notes
`{arch}.attention.q_lora_rank`	`q_lora_rank`	Kimi K2.6: 1536
`{arch}.attention.kv_lora_rank`	`kv_lora_rank`	Kimi K2.6: 512
`{arch}.attention.key_length_mla` (or `.key_length`)	`qk_nope_head_dim` = key_length − rope.dim	Kimi K2.6: 192 − 64 = 128
`{arch}.attention.value_length_mla` (or `.value_length`)	`v_head_dim`	Kimi K2.6: 128
`{arch}.rope.dimension_count`	`qk_rope_head_dim`	Kimi K2.6: 64

For per-head dims the loader prefers the `_mla` variants when present — those carry the pre-absorption (DeepSeek-V3-standard) split that `mla_absorb::absorb` operates on. Kimi K2.6's GGUF exposes both forms (192/128 in `_mla`, 576/512 absorbed); we want the 192/128.

Verification

`cargo test -p larql-models` — 281/281 pass.

Three new tests:

1. `test_kimi_k2_gguf_to_config_json_extracts_mla_fields` — synthesises Kimi K2.6-shaped metadata (q_lora=1536, kv_lora=512, key_length=576+key_length_mla=192, value_length=512+value_length_mla=128, rope.dim=64); checks all MLA fields end up in the HF config, then drives `detect_from_json` and asserts `uses_mla() == true` with the pre-absorption dims.
2. `test_gguf_mla_falls_back_to_non_mla_key_length_when_mla_keys_absent` — older DS-V2 GGUFs that ship only `key_length`/`value_length` still produce the correct split.
3. `test_gguf_mla_fields_absent_for_non_mla_architectures` — llama / qwen / mistral etc. don't emit MLA keys; loader leaves every optional MLA field unset so streaming path keeps its existing behaviour (no regression).

What this unlocks

Combined with the three already-merged PRs (#96 MLA absorption + #103 Q3_K/Q5_K dequant + #133 GGUF-input fix), this PR completes the chain: `larql extract --level inference` works end-to-end on Kimi K2 family GGUFs. Same path works for any DeepSeek-V2/V3 GGUF that exposes the standard MLA metadata.

I plan to extract Kimi K2.6 UD-Q8_K_XL once this lands — happy to share the resulting vindex `index.json` for sanity-check.