Chat routes stop only on the declared EOS; completions stop on every end-of-generation marker (#4661) #4810

Sub-ticket of #4661 (one of its two causes; the other is #4804).

The chat routes (/v1/chat/completions, /api/chat) build their stop set from the declared EOS id alone, while /v1/completions uses completion_stop_tokens: the EOS plus every end-of-generation marker in the vocabulary (<|im_end|>, <|endoftext|>, <|eot_id|>, <|end_of_text|>, <end_of_turn>). A Qwen instruct GGUF declares <|im_end|> as EOS. When a turn ends on <|endoftext|>, chat decodes on to max_tokens while completions stops.

Measured A/B (same tokenizer, /v1/chat/completions, 5 prompts x seeds 1/5/42, CUDA serve):

• Qwen3-Coder-30B-A3B Q4_K_M, "What is 2+2?": base runs to the 256-token limit repeating the prompt (3/3 seeds); with the shared stop set it stops at 8 tokens, "2 + 2 = 4" (3/3). Haiku prompt: 99 -> 24 tokens. The other 3 prompts are unchanged.
• Qwen2.5-0.5B-Instruct f16: unchanged except one row (256 -> 82). Its garbled output is serve: GGUF prompts tokenized greedily, <|im_end|> after text splits into 6 ids (+3 per turn vs llama.cpp) #4804's, not this.

Acceptance:

• chat_stop_tokens returns completion_stop_tokens(tokenizer, Some(eos)), so every chat path (quantized CPU, CUDA dense, CUDA MoE, Qwen3.5, cached, GPU) stops on the same set as /v1/completions; ignore_eos still empties it.
• Unit test chat_stops_on_every_eog_marker_not_just_eos goes through chat_quantized_config (signature unchanged), is RED on the parent commit and GREEN on the fix.
• The 30B "What is 2+2?" chat request stops at 8 tokens for seeds 1, 5 and 42 on the CUDA serve path.
• try_cuda_backend passes the pmat complexity gate (cognitive <= 25) through a behaviour-neutral extraction; cargo test -p aprender-serve --lib is green.