[Core] Carry KV transfer, offload and tracing state across process checkpoints

[galletas1712]

Stacked on #1. Lets prefill and decode workers be checkpointed and restored independently, and reconnect.

• vllm/v1/checkpoint_resources.py: a registry next to the device communicator hooks. Objects that hold state a checkpoint must not contain (network endpoints, NIC registrations, memory shared with other processes) register themselves and implement checkpoint_prepare() / checkpoint_restore(). The engine and each worker run the resources of their own process. New backends plug in without engine changes.
• A resource that cannot be released once started reports restartable = False. Preparation then fails before anything is released.
• A restored process gets a new KV engine id (with its _dp{rank} suffix), passed to workers through checkpoint_restore. Replicas restored from one checkpoint stay distinct, and peers re-handshake instead of reusing cached metadata.
• Waking the KV cache republishes connector handshake metadata before scheduling resumes.
• supports_checkpoint capability for config validation, alongside supports_sleep_mode. NIXL with sleep mode requires --checkpoint-prepare-on-startup. Connectors that cannot follow a KV cache remap release it only for the startup checkpoint.
• NIXL pull: UCX caches network state, such as routes keyed by interface index, for the life of the process. An engine that prepares at startup creates its agent only after restore, and KV registration moves to restore_kv_caches. Checkpointing an existing agent is refused.
• NIXL pull: the side-channel host follows VLLM_NIXL_SIDE_CHANNEL_HOST from the restore environment ([Frontend][Core] Add HTTP checkpoint resume/status lifecycle #1). Without one, the captured value is an interface address of the source pod, so prepare records which interface carries it and restore takes that interface's address in the new pod (same IP family). Names such as DNS names or localhost are kept.
• CPU offloading: the shared host region is unpinned and its pages freed before capture, then pinned again after restore at the same addresses.
• OTEL: otlp_traces_endpoint is the first engine argument a restore may change (checkpoint_restore(env, flags) from [Frontend][Core] Add HTTP checkpoint resume/status lifecycle #1). Every process rebuilds its tracer from it and OTEL_RESOURCE_ATTRIBUTES after restore. OpenTelemetry sets the global tracer provider only once, so vLLM installs a provider that swaps the real one behind it.
• Out of scope: NIXL push mode, the P2P offload tier, and the torch.zeros host buffer fallback.
• Related upstream work: [KV Connector] Support sleep mode with the NIXL pull connector vllm-project/vllm#59635 (NIXL sleep-mode re-registration) and [KV Connector][NIXL] Add restart lifecycle hooks for checkpoint restore vllm-project/vllm#56852 (NIXL restart hooks).