Fix infinite spin-wait and missing cancellation propagation in TaskDeduplicator.

[tjgq]

The test TaskDeduplicatorTest.executeIfNeeded_executeAndCancelLoop_noErrors sporadically hung during the ExecutorService.close() call. This was due to two primary issues:

1. In executeIfNew, if a thread encountered a RefcountedFuture that was already canceled (refcount = 0), it would call Thread.yield() and continue the while(true) loop. It relied on a listener attached to the canceled future to eventually remove it from the map. However, when using virtual threads, many threads could enter this spin loop simultaneously, saturating the underlying carrier threads. This prevented the listener from being scheduled, creating a deadlock where the spinning threads never saw the entry removed.

2. RefcountedFuture was not propagating the cancel() call to its delegate future. This meant that even if all callers canceled their interest in a task, the task would continue to execute in the background, wasting resources and increasing contention.

This PR makes the following changes:

1. executeIfNew and maybeJoinExecution are modified to explicitly call inFlightTasks.remove(key, future) if retain() fails. This ensures the spin loop is broken immediately by the next thread to encounter the canceled future, rather than waiting for an asynchronous listener.

2. RefcountedFuture.cancel now calls delegate.cancel(mayInterruptIfRunning) when the internal reference count drops to zero.

3. Thread.yield() and the associated @SuppressWarnings("ThreadPriorityCheck") are removed, as they are no longer necessary.

Fixes #28302.

[fmeum]

@iancha1992 @meteorcloudy This is another PR for a 9.0.2 release.

[fmeum]

@bazel-io fork 9.1.0

[iancha1992]

@bazel-io fork 9.0.2