xml fallback for llama.cpp models

[areebahmeddd]

summary

closes #2263

• some models served via docker model runner (like qwen3-coder and hermes) ignore the openai function-calling api and instead return tool calls as <tool_call>...</tool_call> text, which caused tools to never execute
• added a fallback to detect and extract these xml blocks after streaming completes and convert them into actual tool calls
• suppressed raw xml blocks from showing up in the tui while streaming so they never appear as visible chat output

ai model disclosure

used vscode copilot (claude sonnet 4.6) to understand the issue context and help generate a fix. changes were self-reviewed and verified via:

task build
task test
task lint