Add DeepSeek-V2 model

[GoodOlClint]

Proposed changes

Adds the DeepSeek-V2 architecture (model_type: "deepseek_v2") to MLXLLM — a port of mlx-lm's deepseek_v2.py. Covers DeepSeek-V2 / V2-Lite, which had no Swift implementation.

It reuses the in-tree DeepSeek-V3 MLA machinery (q_a/q_b + kv_a_proj_with_mqa/kv_b low-rank compression, qk nope/rope split, the shared initializeRope YaRN path) and SwitchGLU MoE with shared experts.

It differs from V3 only in the router (DeepseekV2MoEGate):

• plain softmax scoring (no sigmoid + e_score_correction_bias),
• optional group_limited_greedy group masking by each group's max score (vs V3's top-2-per-group sum),
• an unconditional routed_scaling_factor.

Dense layers below first_k_dense_replace use the standard MLP. sanitize() stacks per-expert gate/up/down projections (weight/scales/biases) into switch_mlp.

Tests (DeepseekV2Tests, no model download): forward-pass logit shape; a deterministic group-limited-greedy routing test (synthetic gate weights → the winning group confines top-k); and sanitize() expert-stacking. Verified via xcodebuild test ... -only-testing:MLXLMTests/DeepseekV2Tests → 3 tests, 0 failures, TEST SUCCEEDED.

Note: validation is structural/unit-level (shape + routing + weight-remap); end-to-end numerical parity against a downloaded DeepSeek-V2 checkpoint has not been run.

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[GoodOlClint]

End-to-end validation (real weights) — caught and fixed a load bug.

Loading mlx-community/DeepSeek-V2-Lite-Chat-4bit-mlx initially failed: V2-Lite sets q_lora_rank: null (direct q_proj, no low-rank q compression), but the config defaulted a null/absent value to 1536, taking the q_a/q_b LoRA path and mismatching the checkpoint's q_proj. Fixed by making qLoraRank optional (nil ⇒ q_proj); added a regression test.

After the fix, forward-pass parity on a fixed token sequence [1..8] vs the Python mlx_lm reference:

	argmax	top5(last)
Python mlx_lm	[185,1,4,5,6,826,4698,575]	[575,2,185,4,746]
this PR (Swift)	[185,1,4,5,6,826,4698,575]	[575,2,185,4,746]

Byte-exact match (MLA + YaRN + group-routed MoE + quantized loading).