fix: drain outbound connections during shutdown

[GrapeBaBa]

Description

Prevent shutdown from finishing while outbound dial jobs or TCP sockets still own referenced native handles.

• mark the connection manager stopped before service teardown
• abort and await active dial jobs
• track outbound TCP sockets and wait for their close events
• avoid resetAndDestroy after a graceful socket shutdown has started, since libuv does not allow mixing uv_shutdown with uv_tcp_close_reset

This was observed while investigating a Lodestar shutdown hang: https://gist.github.com/nflaig/5f41cfc50f38baf5046a034162943dc3

A real failure left a Node.js worker in Environment::CleanupHandles with an inactive but referenced TCP handle. A minimal Node.js reproduction using socket.end followed by socket.resetAndDestroy also hangs with reset EINVAL, while closing the same socket with socket.destroy exits normally.

Verification

• Added real TCP regression coverage for outbound socket close and the reset/graceful-close race.
• Before syncing this branch with the latest main, the affected package lint, doc check, build, and 39 Node.js tests passed.
• The merged head has not been rerun locally. This is a draft PR and CI is expected to validate the latest main integration.
• The patched Lodestar image has completed 23 consecutive mainnet shutdown cycles without a hang so far. The 50-cycle soak is still running.

Notes & open questions

The reset guard preserves resetAndDestroy for ordinary aborts. It uses destroy only when writableEnded shows that graceful shutdown has already begun and a TCP reset can no longer be requested safely.

Change checklist

• I have performed a self-review of my own code
• I have made corresponding changes to the documentation if necessary (this includes comments as well)
• I have added tests that prove my fix is effective or that my feature works

[GrapeBaBa]

Superseded by #3596, whose head points exactly to the source used by the running 50-cycle Lodestar validation image.