[Perf] Early forkserver: fork workers from a preloaded server instead of re-importing

[xenshinu]

Stacked PRs following sgl-project#40201

Motivation:

this PR changes how a worker process is created. Today every worker is a fresh interpreter (spawn) that re-imports the sglang/torch/kernel stack, 7-9 s per worker, and a third time on the DP path where the controller re-imports before it spawns the schedulers. Python's forkserver start method creates one small helper process and forks every child from it; if that helper has imported the worker modules once ("preload"), a forked worker inherits them and imports nothing. The launcher itself cannot be the fork parent: it has threads and, through sgl_kernel, a CUDA context.

Modifications:

SGLANG_EARLY_FORKSERVER=1 starts the multiprocessing forkserver at CLI entry with the scheduler, DP controller and detokenizer modules preloaded; workers fork from it. The child re-applies the launcher's environment and stdio, re-derives the HF offline flags, reports the launcher as its parent, keeps its sglang::* process title, and nested Process() calls reuse the same forkserver. Import-time GPU probes of third-party packages are answered from NVML inside the forkserver; an attribute NVML cannot supply raises a named error. --enable-memory-saver falls back to spawn with a warning. New Envs: SGLANG_EARLY_FORKSERVER (off), SGLANG_MP_START_METHOD (set by start_early()), internal SGLANG_FORKSERVER_ADDRESS/PID. Default behaviour unchanged (engine.py: the start method is read from Envs, default spawn).

Accuracy:

greedy output identical with the flag on and off (Qwen3-1.7B and Qwen3-30B-A3B, 128- and 64-prompt batches). Speed: single-GPU Qwen3-30B-A3B time to /health 47.3 s -> about 43 s with the forkserver alone (the rest of the gain needs PR 3). Tests: entrypoints/test_early_forkserver.py (env merge, offline flags, NVML stand-in, forked-worker title regression, switches).