Upstream: contribute Gemma 4 reasoning format to llama.cpp

[ArtemisAI]

Context

strip-gemma exists because llama-server has no working --reasoning-format option for Gemma 4:

• --reasoning-format deepseek → content field empty, everything in reasoning_content
• --reasoning-format none → raw <|channel>thought\n...<channel|> tags in content
• --reasoning-format auto → doesn't recognize Gemma 4's channel tags

Proposed upstream fix

Teach --reasoning-format auto in llama-server to recognize Gemma 4's <|channel>thought\n...<channel|> tag pattern and split it into reasoning_content vs content, the same way it already handles DeepSeek's <think>...</think> tags.

Target files

• common/chat.cpp — reasoning format parsing/extraction
• common/common.h — enum common_reasoning_format (note: comment says "do not extend unless you absolutely have to" — prefer making auto smarter over adding a new enum value)

Related upstream issues

• Discussion #21338 — Can't disable thinking in gemma4
• Discussion #15333 — How should clients handle thinking blocks?
• Issue #20196 — Disabling reasoning doesn't work on certain models
• PR #21418 — Gemma 4 specialized parser (merged, but doesn't fix reasoning format)

Affected communities (same problem, different stacks)

• KoboldCpp #2092 — 19 comments, Gemma4 think tags not detected
• vLLM #38855 — 7 comments, reasoning parser fails
• Open WebUI #23357 — Gemma4 changes required

Acceptance criteria

• --reasoning-format auto correctly splits Gemma 4 channel tags in non-streaming responses
• --reasoning-format auto correctly splits Gemma 4 channel tags in streaming (SSE) responses
• Truncated thinking (no closing tag) handled gracefully
• No regressions for DeepSeek or other model reasoning formats
• PR submitted to ggml-org/llama.cpp