Fix streaming pipeline: reasoning parser support, context limits, and…

[bruceyxli]

… query reformulation

• base_handler: robust reasoning_content reading via model_extra fallback ('reasoning'/'reasoning_content'); fix flush block to recompute channels from accumulated text; update previous_channels after flush
• model_call: reduce max_tokens to 2000 to stay within 10000 context limit; add thinking_budget: 512 to sampling params
• parser: strip tag from analysis channel in streaming/incomplete case for consistency; add all partial suffixes to incomplete_patterns; handle vLLM stripping opening token
• reformulation: singleton AsyncOpenAI client to fix connection pool leak; remove garbage strings from prompt; add timeout, reasoning fallback, and user_message fallback
• start_vllm_servers: update model path, GPU memory utilizations