… support

- Add _sanitize_generate_config to MLXVisionModel to handle stop/stop_token_ids from model family
- Initialize reasoning parser in MLXVisionModel.load() for hybrid models
- Pass reasoning_parser to completion methods for proper content parsing
- Pass tokenizer to get_full_context for correct jinja template rendering
- Set enable_thinking=False as default for MLX models
- Enable stream by default for multimodal interface in Gradio UI

…abled

- Fix _force_virtualenv_engine_params: only include engines with matching specs,
  previously fallback to all specs when matched_specs is empty
- Add Gemma4ForConditionalGeneration to VLLM_SUPPORTED_MULTI_MODEL_LIST for vllm >= 0.19.0

- VLLMMultiModel: fall back to tokenizer's chat_template when model_family.chat_template is empty
- SGLANGVisionModel: same logic as VLLMMultiModel, align with MLX core behavior
- SGLANGModel: use empty string instead of asserting when chat_template is empty

When chat_template is empty, engines will:
1. Get tokenizer and retrieve its chat_template attribute
2. If still empty, raise ValueError
3. Pass tokenizer to get_full_context for proper template application

This aligns the behavior across vllm, sglang, and mlx engines.

When thinking mode is enabled (enable_thinking=True), special tokens are
needed for the thinking/reasoning format. Set skip_special_tokens=False
in this case to preserve the special tokens.