feat(models): add Kimi K2 + K2.7 (text-only) support

[faresobeid]

Summary

• Kimi K2 reuses DeepSeek-V3's architecture verbatim (MLA + MoE hyperparameters, weight-key naming — Moonshot's own checkpoints ship transformers' DeepseekV3ForCausalLM as their auto_map target). KimiK2Config subclasses DeepseekV3Config directly rather than redefining its ~20 MLA/MoE fields.
• New layers/mla.py: plain (dense, non-absorbed) Multi-head Latent Attention, shared infrastructure any MLA family can build on.
• MoE HF↔prime conversion is reused as-is from glm4_moe — the per-expert gate/up/down + shared-experts + e_score_correction_bias layout matches exactly, no adaptation needed.
• K2.7 (and K2.5/K2.6) ship as a multimodal wrapper around this same text backbone (language_model.-prefixed weights + a vision tower). Loading one strips the prefix and drops the vision tower entirely — text-only, no VLM/multimodal support.
• Builds on a resurrected/adapted version of @samsja's earlier sami/kimi-k2 draft (add kimi k2  #1761, closed unmerged) — same core design (subclass DeepseekV3Config, reuse GLM's MoE conversion), updated for current main (renamed kimi_k25→kimi_k2, fa4→flash_attention_4, get_cu_seqlens_from_position_ids, MoEArgs.fp8), plus the K2.7 text-extraction addition and a dedup pass on the two MLA attention variants.

Verification

• scripts/mini_moe.py --arch kimi_k2: HF↔prime roundtrip logits match to 1e-6, using transformers' own DeepseekV3ForCausalLM as the reference (real Kimi checkpoints use this exact class, so this validates the modeling code faithfully reimplements the same math).
• Full trainer unit test suite: no new failures from this change (confirmed the same pre-existing failures reproduce identically on a clean, unmodified checkout on this GPU node — unrelated to this PR).
• No real Kimi checkpoint was available to validate end-to-end against.
• No inference-side changes: vLLM already serves real Kimi checkpoints natively, since their config.json declares architectures: ["DeepseekV3ForCausalLM"] and routes straight into vLLM's existing DeepSeek-V3 support.

🤖 Generated with Claude Code