Fix libdispatch over-resume crash in DispatchTimerSource.deinit

[romansorochak]

Problem

DispatchTimerSource.deinit resumes the timer unconditionally:

deinit {
    timer.setEventHandler(handler: nil)
    timer.cancel()
    // balance called of cancel/resume to avoid crashes
    timer.resume()
}

That resume is necessary when the source is suspended — libdispatch traps when a
suspended source is released. But it is an over-resume when the source is still
activated, and libdispatch traps on that just the same.

The activated case is reached on a normal path in this library. Retrier schedules a
one-shot timer and calls timeoutTimer.activate() in internalRetry(); the timer stays
in .activated until someone calls cancel(). Whenever a failing stream is torn down
with a retry pending, the Retrier — and with it the DispatchTimerSource — is
deallocated while the source is still running, and the process aborts.

Crash signature:

Thread 0 Crashed:
0  libdispatch.dylib  _dispatch_source_dispose
...
   xctest at DispatchTimerSource.deinit
libdispatch: BUG IN CLIENT OF LIBDISPATCH (Abort Cause 105553162484480)

For context on how often this path is hit: in a radio app built on AudioStreaming this
was 62% of all iOS crash events, affecting 1.3% of weekly active users — dead and
flaky stream URLs put Retrier to work constantly, and every teardown with a pending
retry was a coin flip.

Fix

Make the balancing resume conditional on the source actually being suspended, and take a
lock around the check-and-change in activate() / suspend() so the suspend count cannot
be torn when those are called from different queues. state becomes a lock-protected
read-only property; isRunning is unchanged.

Tests

Two regression tests in DispatchTimerSourceTests:

• test_DispatchTimerSource_Can_Be_Deallocated_While_Activated — deallocates an activated
  source. Reaching the end of the test is the assertion: against the current main this
  test does not fail, it takes the whole test runner down at DispatchTimerSource.deinit.
• test_DispatchTimerSource_Repeated_Activate_And_Suspend_Stay_Balanced — repeated
  activate() / suspend() calls move the suspend count exactly once each, including the
  no-op paths.

Notes

The change is confined to DispatchTimerSource; no public API changes. We have been
running this patch in production on a pinned fork revision, and would rather have it
upstream than carry the fork.