common : fix Hermes tool-call parser leaking wrapper XML

[laundrevity]

This has particularly been an issue when trying to use the qwen-code CLI in conjunction with llama.cpp - it essentially only works with vLLM for Qwen3-Coder models.

changes:

• strip <tool_call>/<tool> wrappers when parsing Hermes/Qwen function-tag tool calls so they no longer leak into assistant content
• add parser helper to drop wrapper-only preludes before consuming <function=...> blocks
• cover nested <tool_call><function=...> cases with a regression in test-chat.cpp