fix(distserve): coalesce PD connection warmup to prevent Proxy OOM

[gokay-ai]

Motivation

Fixes #4972.

/distserve/connection_warmup gathered connect() for every Prefill×Decode pair with no in-flight coalescing or waiter timeout. Slow or stuck handshakes piled waiting_conn items and wait_for_conn tasks until the Proxy was OOM-killed (exit 137).

Modification

• Per-link single-flight in PDConnectionPool: later connect() calls join the existing Connecting event instead of enqueueing another handshake or wait_for_conn task.
• Bound connect() waiters with LMDEPLOY_PD_CONN_WAIT_TIMEOUT (default 60s; 0 restores unlimited wait).
• Cap waiting_conn as a safety net.
• Make warmup idempotent: concurrent /distserve/connection_warmup callers share one in-flight gather, with a semaphore limiting concurrent handshakes.
• Add unit tests for coalescing and waiter timeout.

BC-breaking (Optional)

No API break. DistServe-only. Operators who need unbounded PD handshake waits can set LMDEPLOY_PD_CONN_WAIT_TIMEOUT=0.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[RunningLeon]

may put env variable in https://github.com/InternLM/lmdeploy/blob/main/lmdeploy/pytorch/envs.py

[gokay-ai]

Moved LMDEPLOY_PD_CONN_WAIT_TIMEOUT into lmdeploy/pytorch/envs.py so get_all_envs() registers it.