Let the poll server restart after CleanupSRT

[jeoliva]

Fixes a regression I introduced in the previous commit, caught by testing a start/stop/start cycle that nothing in the suite covered.

The bug

CleanupSRT stops the poll server and releases its epoll. But the poll server was created behind a sync.Once, so it could never be replaced. The next non-blocking socket tried to register against the released epoll:

panic: ERROR ADDING FD TO EPOLL
  pollserver.go:82 -> poll.go:109 -> srtgo.go:122

This sequence worked at 9576f32 and panicked after:

InitSRT(); sock := NewSrtSocket(...); sock.Close()
CleanupSRT()
InitSRT(); sock2 := NewSrtSocket(...)   // panic

Start/stop/start is a reasonable thing for a consumer to do — and the whole point of CleanupSRT being exported.

The fix

Replace the sync.Once with a nil-checked, mutex-guarded pointer. pollServerShutdown clears it; the next socket constructs a fresh poll server with a fresh epoll. shutdown() itself stays idempotent via its own stopOnce.

Verification

• The new TestPollServerRestartsAfterCleanup panics without this change and passes with it.
• The exit-teardown fix is unaffected: full suite under -race, no segfaults, 10/10 on macOS arm64 and 10/10 on Linux amd64.

CI did not catch this because no test exercised a restart. That gap is now closed.

🤖 Generated with Claude Code