Rework router worker sync around job ID labels

[un-def]

The sync matched SMG router workers to jobs by URL and removed every worker that wasn't a ready dstack worker, so users couldn't register workers of their own. It also probed every worker on each sync and deregistered the ones that failed, so a flaky connection from the server made a worker flap even when the router, which reaches it within the cluster, had no problem with it. Probes ran one at a time, and only for the connection mode and runtime type of the workers already registered, which broke services that mix them or switch between them in a rolling deployment.

Now each worker that dstack registers is labeled with its job ID (dstack.ai/job-id), and the sync works in terms of jobs:

• A worker is probed only while some router lacks it. Once it's registered, its health is left to the router, which checks it anyway. A worker that is truly unreachable has its job terminated, which deregisters it.
• A worker is removed once its job is no longer running. Workers that dstack can't map to a job are left alone, so users can register their own.
• Workers registered before the label existed are mapped to jobs by address. The ones whose job has already stopped are no longer removed: the router marks them unhealthy, and a rolling deployment of the router drops them.
• Unregistered workers are probed for HTTP, SGLang gRPC and vLLM gRPC at once, with no hints from the registered ones.
• Routers and workers are processed in parallel, up to 8 replicas at a time. An unexpected error in one of them is logged and no longer aborts the sync for the others.

Also:

• gather_async() is added, and gather_map_async() is built on it. Both accept max_concurrency, and cancel the remaining awaitables when one fails unless return_exceptions is set.
• Router responses are validated with pydantic.