Add GLM-4 reasoning parser and fix think tag / prefix cache bugs

[janhilgard]

Summary

• Add Glm4ReasoningParser (--reasoning-parser glm4) for GLM-4 models (GLM-4.5-Air, GLM-4.7) that use <think>...</think> tags but don't inject <think> in the prompt. Unlike Qwen3 where no tags = truncated reasoning, GLM-4 no tags = normal content. Streaming override emits pre-think deltas as content.
• Fix duplicate </think> tag in BaseThinkingReasoningParser: some models generate <think></think></think> with an extra closing tag that leaked into content.
• Fix reasoning extraction with tool calls: always parse reasoning from original output.text so <think> content is preserved even when tool calls are present.
• Remove <tool_call> from SPECIAL_TOKENS_PATTERN: these are structural tags used by GLM-4's tool calling format (--tool-call-parser glm47) and must not be stripped.
• Fix prefix cache corruption on exact match: trim stored cache by 1 in _prompt_cache_save so the last prompt token is reprocessed at the correct position on cache hit.

Test plan

• Basic text generation (reasoning + content correctly separated)
• Tool calling with reasoning (tool_calls + reasoning_content in response)
• Streaming (reasoning chunks → content chunks transition)
• Prefix cache exact hit (identical output on repeated prompts)
• KV cache quantization Q8 (no degradation)
• Verified Qwen3.5-122B (MLLM path) unaffected by scheduler.py fix

Tested with GLM-4.5-Air (106B/12B MoE, mixed 6/8-bit) on Apple M3 Ultra, ~37 tok/s.

🤖 Generated with Claude Code