Ensure thread is released on completion

[awilfox]

This prevents leaking Fiber objects when running timeouts in nested threads.

Tests pass locally on Darwin arm64 (3.4), Linux ppc64 (3.4), and Windows x86_64 (4.0).

[OskarEichler]

I verified this retention fix against the current 0.6.1 code and found one adjacent payload-retention path: after a request completes, the retained Request still holds its exception class and message. With a 1 MiB message, the existing consumer retained the payload in 10/10 runs on Ruby 4.0.6 and 3.2.11.

I prepared a one-line follow-up directly on this PR head: OskarEichler@67e0435

It clears @exception_class and @message under the same mutex alongside @thread. The focused payload model releases both fields in 20/20 runs on each Ruby; the combined suite passes 284 tests / 636 assertions with zero failures. No repository tests were changed. I kept this as a follow-up on #114 to preserve the original contribution and avoid a competing PR.

[eregon]

Sorry for late review.

This seems redundant and unnecessary, because the Request object will GC soon after finished is called, and the timeout thread has @queue.pop'd it, as then nothing will refer to it.
If there indeed an issue, please provide a reproduction.

This prevents leaking Fiber objects when running timeouts in nested threads.

How are Fibers involved here?
What do you mean by "nested threads"?