fix(models/kimi-vl): parse real Kimi-VL text config (nullable field)

[inureyes]

Follow-up hardening for #525: the merged Kimi-VL port fails to load its real checkpoint.

Problem

mlxcel generate -m /path/to/kimi-vl-a3b-thinking-4bit ... aborts during config parsing:

Failed to parse Kimi-VL text config: invalid type: null, expected usize

The real text_config (a DeepSeek-V3-style block) sets q_lora_rank to null. The Moonlight-16B backbone that Kimi-VL uses projects the query directly rather than using a LoRA-compressed query, so upstream stores q_lora_rank: null. DeepSeekV3Config declared the field as a bare usize, and serde rejects null for usize.

Change

• DeepSeekV3Config::q_lora_rank becomes Option<usize> with #[serde(default)], so both an explicit null and an absent key parse.
• DeepSeekV3Attention threads the None case through the model build: a null q_lora_rank loads a direct self_attn.q_proj weight; a numeric value keeps the LoRA-style q_a_proj -> q_a_layernorm -> q_b_proj chain. This mirrors the existing DeepSeek-V2 MLA path. The real checkpoint's safetensors index confirms it ships self_attn.q_proj and no q_a_proj / q_b_proj.

Shared-path consistency

DeepSeekV3Config and DeepSeekV3Attention are reused by the standalone DeepSeek-V3 loader, the pipeline stage executor, the Kimi-VL loader, and Youtu-VL:

• Kimi-VL and any Moonlight-style backbone (q_lora_rank: null) now load via the direct q_proj path.
• Youtu-VL wraps its required numeric rank in Some, so it keeps taking the LoRA branch (no behavior change).
• Genuine DeepSeek-V3 (numeric q_lora_rank) is unchanged.
• DeepSeek-V3.2 and GLM-MoE-DSA keep their own separate config structs and are not affected.

Tests

• src/models/deepseek_v3.rs: parses_null_q_lora_rank_config (real config shape with q_lora_rank: null) and parses_numeric_and_absent_q_lora_rank.
• src/loading/vlm_kimi_vl.rs: parses_real_kimi_vl_text_config_with_null_q_lora_rank exercises the loader's parse_required_vlm_subconfig call with the real config shape.

Verification

• cargo test --lib kimi_vl (16 passed)
• cargo test --lib deepseek_v3 (15 passed)
• cargo check --lib --tests
• cargo clippy --lib --tests -- -D warnings
• cargo fmt

Real-model GPU load is deferred to the orchestrator.