common : fix gpt-oss Jinja error with content and thinking on tool-call messages

[abhijitb11]

Summary

• Fix Jinja template error "Cannot pass both content and thinking in an assistant message with tool calls" when using gpt-oss models (e.g. ggml-org/gpt-oss-120b-GGUF) with --reasoning-format auto in multi-turn tool-call conversations
• Erase content from the adjusted message after copying reasoning_content to thinking in common_chat_params_init_gpt_oss()
• Regression from common : move gpt-oss reasoning processing to init params #16937

Fixes #19701

Details

When a client sends back conversation history containing assistant messages with content, reasoning_content, and tool_calls, the gpt-oss input processing copies reasoning_content → thinking but leaves content in place. The gpt-oss Jinja template forbids having both content and thinking on tool-call messages (they render to the same <|channel|>analysis slot), causing an HTTP 500 error.

The fix is a single line: adjusted_message.erase("content") after setting thinking.

Test plan

• Served ggml-org/gpt-oss-120b-GGUF with --reasoning-format auto --jinja
• Sent multi-turn chat completions with tool calls — confirmed no more 500 errors
• First-turn requests continue to work as before

🤖 Generated with Claude Code

[abhijitb11]

Closing — rewriting to comply with the project's AI usage and contributing guidelines.