feat: add InternLM2/2.5 (InternLM2ForCausalLM) loader to aprender::rosetta

[noahgift]

Context

The cookbook architecture-demos spec tracks InternLM2.5 as status: blocked. Issue #367 ("feat: add InternLM2.5 architecture support for inference", closed 2026-04-13) was deferred ("Low — skip in QA campaign for now"). This is the re-open: the cookbook now blocks 1 family-smoke recipe + apr-model-qa-playbook playbooks until the loader lands.

Family

• Name: internlm2_5
• Vendor: InternLM
• HF architectures: InternLM2ForCausalLM
• HF pattern: internlm/internlm2_5-*
• Reference checkpoints: internlm/internlm2_5-1_8b-chat, internlm/internlm2_5-7b-chat, internlm/internlm2_5-20b-chat

Known gating constraint (from #367)

InternLM2 uses non-standard tensor naming: attention.wqkv (fused QKV) instead of self_attn.q_proj.weight. The loader needs to recognize this naming variant and either remap or generate fused-QKV-aware kernels.

Acceptance criteria

• contracts/model-families/internlm2_5.yaml exists with size_variants for 1_8b, 7b, 20b
• Loader handles attention.wqkv fused-QKV tensor naming
• SentencePiece tokenizer embeds correctly (already fixed in apr convert: APR file missing embedded tokenizer for SentencePiece models (internlm2.5) #366; verify it still works for internlm)
• Inference smoke pass against internlm2_5-1_8b-chat resolves the missing-tensor failure documented in feat: add InternLM2.5 architecture support for inference #367

Unblock impact

• Cookbook manifest flips from blocked to certified
• 1.8b checkpoint is small enough for sub-1B dim-smoke regression coverage

Cookbook reference

• manifest.yaml entry — name: internlm2_5 block
• Closed predecessor: aprender#367