feat(rollout-gateway): make renderer derender stages pluggable

[Linbo-Liu]

Issue #, if available:

Description of changes:
HfTemplateRenderer derendered model output with one hardcoded path: split reasoning on </think>, then match tool calls with a regex for the <tool_call><function=...> XML format. Any model whose tool-call format differs silently produced zero tool calls.

For example, Qwen3-0.6B has parsing format:

  <tool_call>
  {"name": <function-name>, "arguments": <args-json-object>}
  </tool_call>

while the parser can only handle

  <tool_call>\s*<function=NAME>...<parameter=K>V</parameter>...</function>\s*</tool_call>

Stage	Signature	Default
reasoning_parser	raw_output -> (reasoning, body_text)	split on </think>
tool_parser	(body_text, tools_schema) -> (text, tool_uses, ill_formed)	<tool_call><function=...> regex

The stages run in sequence — reasoning first, tool calls on what remains — matching the order the inference engines use. Overriding one leaves the other on its default, so a model needing engine-grade tool parsing but emitting plain </think> reasoning supplies only tool_parser. The tool stage is skipped entirely when a request carries no tools schema.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[luyuzhe111]

wondering why we dropped the async here?

[Linbo-Liu]

We have a threaded server that can start rollout gateway even in a sync function call (both slime and verl invoke the gateway in sync caller), this integration test is to test whether threaded server can work in a sync caller.

[lyzustc]

Is it possible that we build a mapping from supported model family to parser functions?

[Linbo-Liu]

Do we want to implement the parser functions or import from inference engine? Our tenet is to have 0 dependency on vllm/sglang.

[lyzustc]

I am ok to use some lightweight packages like renderers (https://github.com/PrimeIntellect-ai/renderers), or implement and maintain ourselves. Anyway, I think it may improve user experience if we can automatically match the correct parser functions according to the model family used by users?