[tx] Tinker CI + bugfixes

[hao-aaron]

This PR introduces a tinker ci, closing #1553. It uses tinker_cookbook.recipes.math_rl.train. Currently the REWARD_MIN_VALUE is set to 0, but will be tightened once we have a couple of runs. Added both colocated and fully async tests.

https://wandb.ai/sky-posttraining-uc-berkeley/gsm8k_tinker_ci?nw=nwuserahao
https://wandb.ai/sky-posttraining-uc-berkeley/gsm8k_tinker_fully_async_ci?nw=nwuserahao

While testing out the tinker CI, found a port collision bug.

async def _run_server(self) -> None:
    """Internal method to run the HTTP server."""
    # Release the port reservation right before vLLM rebinds.
    if self._port_reservation is not None:
        self._port_reservation.close()                          # ← step A
        self._port_reservation = None
    ...
    sock_addr = (self._cli_args.host, self._cli_args.port)
    sock = create_server_socket(sock_addr)

There exists a window of time where we free port so vllm can take it, but other actors can also. Fixed this by allocating a specific, non-overlapping port range for each actor.

Additionally, changed the disk reloading workaround #1606 created here to more closely mirror the upstream change vllm-project/vllm#41482, adding a new custom endpoint with the load_inplace override.

[gemini-code-assist]

Code Review

This pull request introduces a custom LoRA loading endpoint /skyrl/v1/load_lora_adapter to bypass upstream vLLM bugs and implements a port striding mechanism (SERVER_PORT_STRIDE) to prevent port conflicts between inference servers. It also adds CI configuration and E2E test scripts for the Tinker backend. Feedback includes adding a check for LoRA support in the new endpoint to avoid potential attribute errors and ensuring consistent return formats in the inference client.

[gemini-code-assist]

The _skyrl_load_lora_adapter endpoint assumes that LoRA is enabled on the server. If the server is started without LoRA support (e.g., --enable-lora is not passed to vLLM), models.lora_requests and models.lora_resolver_lock may not be initialized, leading to an AttributeError or TypeError. A check should be added to ensure LoRA is enabled before accessing these attributes.

            if not hasattr(models, "lora_requests") or models.lora_requests is None:
                raise HTTPException(
                    status_code=400,
                    detail="LoRA is not enabled on this server.",
                )

            async with models.lora_resolver_lock[lora_name]:

[gemini-code-assist]

The return value of _load_on_server is inconsistent with other control plane methods in this class (e.g., _call_server at line 857), which return a dictionary containing both the HTTP status and the response body. This inconsistency may cause issues for callers expecting the standard {url: {"status": ..., "body": ...}} format.

Suggested change

	return server_url, await resp.json()
	return server_url, {"status": resp.status, "body": await resp.json()}

[SumanthRH]

I would prefer if the default tinker test is non-colocated, 2 GPUs for trainer and 2 for inference.

[SumanthRH]

Actually can you just add a separate non-colocated E2E test?

[hao-aaron]

added