[V1]: Isolate IPC endpoints per server run

[Ratish1]

Motivation

Omni V1 currently derives default IPC endpoints from endpoints.base_path / config.name. When two V1 servers are launched with the same model name and default IPC settings, they can bind/connect to the same ZMQ socket paths and leak control-plane messages across server instances.

This applies the same per-server IPC namespace invariant used for the V0 runtime isolation fix in #263, adapted to the V1 compiler and StageProcessSpec startup path.

Root Cause

PipelineConfig.name defaults to the model path. Two Qwen3-Omni V1 servers launched with the same model therefore allocate identical stable endpoints under the same model-name directory.

Because V1 uses ZMQ for Coordinator-to-Stage control-plane traffic, sharing these paths can let one server receive another server's stage messages or completions.

Modifications

• Add a managed IpcRuntimeDir lifecycle in sglang_omni_v1.config.compiler, matching the V0 per-run IPC directory ownership model from [Server]: Isolate IPC Endpoints Per Server Run #263.
• Allocate IPC endpoints under a unique runtime directory created with tempfile.mkdtemp(...) instead of the stable base_path / config.name path.
• Split V1 compilation into prepare_pipeline_runtime(...) and compile_pipeline_core(...) so callers that own server lifecycle also own IPC cleanup.
• Keep V1 multi-process startup V1-native: MultiProcessPipelineRunner prepares one endpoint map in the main process and passes it into StageProcessSpec construction, so subprocesses still do not recompile endpoint state.
• Use the managed compile path in the V1 single-process server launcher and close the runtime directory during shutdown/failure cleanup.
• Add V1 IPC regression coverage for same-model namespace uniqueness, non-IPC no-op behavior, unmanaged IPC rejection, compile failure cleanup, caller-owned cleanup, successful compile ownership, multi-process cleanup, and single-process launcher cleanup.

Accuracy Tests

• Unit tests: PASS
• V1 runtime endpoint preflight: PASS
  • observed ipc:///tmp/sglang_omni_v1/qwen-qwen3-omni-30b-a3b-instruct-9ta1b5i2/stage_preprocessing.sock
  • observed ipc:///tmp/sglang_omni_v1/qwen-qwen3-omni-30b-a3b-instruct-9ta1b5i2/stage_thinker.sock
  • observed ipc:///tmp/sglang_omni_v1/qwen-qwen3-omni-30b-a3b-instruct-9ta1b5i2/completion.sock
• Two Qwen3-Omni V1 text servers launched side by side with default IPC settings: PASS
  • server on port 8101 submitted to ipc:///tmp/sglang_omni_v1/qwen-qwen3-omni-30b-a3b-instruct-y3dtxwhb/stage_preprocessing.sock
  • server on port 8102 submitted to ipc:///tmp/sglang_omni_v1/qwen-qwen3-omni-30b-a3b-instruct-1w6lk5as/stage_preprocessing.sock
• Sent concurrent simple curl chat-completion requests to both servers and received HTTP 200 responses from both: PASS

Speed Tests and Profiling

Not run. This changes server startup endpoint allocation and shutdown cleanup only; it does not touch request scheduling, model execution, relay tensor transfer, or decode hot paths.

[zhaochenyang20]

Please rebase and rerun CI after #403 is merged