[deps] Upgrade vllm-router and fix 0.1.15 startup

[bvolpato]

Summary

Replace temporary vllm-router==0.1.14.post1 custom wheels with official vllm-router==0.1.15 across SkyRL training, FSDP, and Megatron extras.

Version 0.1.15 includes the custom wheel's load-aware routing fix, publishes x86_64 and aarch64 wheels, and removes the old pd_disaggregation RouterArgs field. SkyRL's fallback access to that removed field crashed regular, non-PD startup after the router became healthy.

Changes

• Pin official vllm-router==0.1.15 and refresh the lockfile.
• Use the 0.1.15 vllm_pd_disaggregation field directly.
• Register SkyRL's typed /inference/v1/generate endpoint through router configuration.
• Cover regular and prefill/decode startup arguments.

Testing

• Router tests (10 passed)
• Direct construction of regular cache-aware and prefill/decode RouterArgs
• Direct import of official vllm-router==0.1.15
• uv lock --check
• Ruff, Black, Gitleaks, and git diff --check
• Lockfile contains official manylinux x86_64 and aarch64 wheels

[gemini-code-assist]

Code Review

This pull request updates the vllm-router dependency to the official 0.1.15 release across the skyrl-train, fsdp, and megatron dependency groups in pyproject.toml. It also removes the temporary custom wheel definitions that were previously used as a workaround. There are no review comments, and I have no feedback to provide.

[SumanthRH]

@bvolpato the custom wheel we're using also has our custom load aware router policy sticky_least_loaded: vllm-project/router@main...SumanthRH:router:sticky_least_loaded

So we can't simply upgrade to vllm-router 0.1.15 directly .

Let me make an issue to track the upgrade though - we should move to a custom wheel on top of 0.1.15 for now.

[SumanthRH]

Created an issue to track #2070

[bvolpato]

Makes sense! Thanks @SumanthRH