Continuous batching: model support tracker #50352

Tracking which model families continuous batching (generate_batch, ContinuousBatchingManager) runs. The engine takes its attention types from layer_types (group_layers_by_attn_type) and sizes its paged cache from num_key_value_heads and head_dim; anything outside that is rejected at init or not representable.

• decoder-only models with full_attention and/or sliding_attention layers (dense and MoE), with paged|eager, sdpa, flash attention 2/3; tensor parallelism
• gated delta net hybrids: Qwen3-Next, Qwen3.5, Qwen3.5-MoE Continuous batching for linear attention layers #50350
• OLMo-Hybrid: own copy of the gated delta net forward (declare its states, add the packed-row branch: the smallest one)
• Kimi Linear, GLM5-next: KDA (KimiLinearAttention, Glm5NextTextLinearAttention)
• MiniMax: MiniMaxLightningAttention
• Bailing-Hybrid: BailingHybridDeltaAttention
• Mamba2 mixers: Bamba, Falcon-H1, NemotronH, Granite-MoE-Hybrid, Jamba, Zamba, Zamba2, Inkling, Nemotron-3.5-Transcribe (conv state + SSM state, causal_conv1d and mamba_ssm with cu_seqlens)
• mamba layers (Mamba, Jamba) and conv layers (LFM2): same kind of per-request state
• Qwen4-Exp: indexed_attention (QSA indexer) and the per-layer n-gram embeddings Run Qwen4-Exp under continuous batching #50359
• indexed_attention with MLA: DeepSeek-V3.2, GLM-5, Mistral Large 4, HY-V4, AXK2 Run multi-head latent attention models under continuous batching #50365
• chunked_attention (Llama 4)
• MLA: DeepSeek V2/V3, Kimi K2, GLM-4.7-Flash, MiniCPM3, Mistral 4, LongCat-Flash, Youtu, AXK1 Run multi-head latent attention models under continuous batching #50365
• DeepSeek V4: compressed sparse and heavily compressed attention (one compressed entry per window of tokens, the compressor's open window as a per-request state, sliding keys)
• multimodal prompts: generate_batch takes token ids only, no pixel inputs or inputs_embeds, so VLMs run text-only
• encoder-decoder models

Groundwork that makes the per-family items mechanical: a varlen scan with per-sequence states for Mamba2, like causal_conv1d_varlen for the convolution. The cache takes the rows and states each layer declares, by name (cache_rows, cache_states, #50356), so a new family is that declaration in its modules plus the packed-row branch of its forward.