Signed-off-by: Oleh Dokuka <odokuka@vmware.com>

This commit makes sure that the picked worker is released (calling
dispose() method) if a RejectedExecutionException is raised when the
BoundedElasticScheduler tries to submit a task to its picked worker.

Fixes #3182.

Co-authored-by: Stanislav Kotik <STAlkerS-05@yandex.ru>
Co-authored-by: Kuznetsov Stanislav <SMikhKuznetsov@sberbank.ru>

This PR improves the japicmp integration in the build in several ways:

1) Add a finalizedBy task that prints a filtered report
The basic text report uses * and ! as markers for binary and source
incompatible changes respectively. Additionally, a MODIFIED class is
marked with ***. This task attempts at pinpointing the incompatible
changes only in the report, and printing these lines. This is called out
in the ci job when prepare step fails fast.

2) Ensure sourceModified changes are considered
See reactor/reactor#722.
If a change is binary compatible but not source compatible, using
`onlyBinaryCompatibleModified = true` will mask the issue. This
change uses `onlyModified = true` instead. This is slightly more
verbose, but this is alleviated by (1).

3) Bump to japicmp plugin v0.4.1, exclude NEW_DEFAULT_METHOD as a whole
The new `compatibilityChangeExcludes` parameter allows us to ignore all
occurrences of a default method added to an interface, instead of
having to exclude each such method one by one.

…lux.cache(int) (#3200)

Supporting the caching of only terminal signals in case of Flux.replay(int) and Flux.cache(int) operators.

As it currently stands, these operators, when provided 0 as the argument, resort to behaving like Flux.publish() which differs in the way termination signals are handled by the > 0 cases. This change still uses the FluxPublish class to implement the logic, however with a few minor changes to its implementation.
First of all, FluxPublish resets itself after source termination to be able to connect() again. FluxReplay on the other hand, does not reset itself in the case of only buffering signals without timeout. Therefore, for the behaviour of caching the terminals, FluxPublish does not reset itself and replays the terminal when it receives a late subscription.
For cases with expiry (TTL arguments), FluxReplay does indeed reset itself. Therefore, the behaviour for the time constrained values remains as before, using FluxPublish implementation for the 0 history case, but without caching terminals, while not honouring the TTL. This case can be later implemented if needed.

Fixes #3164

This fixes a situation where the first non-matching onNext signal could
be skipped if the thenConsumeWhile was followed by a `TaskEvent` step.

Fixes #3121.

This prevents BlockHound from flagging (short-lived) blocking calls that cannot
 always be avoided when a `dispose` call results in the removal of a task
 from the worker-backed queue when access to said queue is currently held
 by a guard.

Fixes #3210.

…3218)

This commit updates the following dependencies:

- bytebuddy 1.12.14 -> 1.12.17, supersedes and closes #3186

It also updates the following build dependencies:

- plugin artifactory 4.29.0 -> 4.29.1, supersedes and closes #3207
- plugin jcstress 0.8.13 -> 0.8.14, supersedes and closes #3171
- plugin download 5.1.2 -> 5.2.1, supersedes and closes #3188
- plugin spotless 6.10.0 -> 6.11.0, supersedes and closes #3193

And also updates the following test dependencies:

- JUnit 5.9.0 -> 5.9.1, supersedes and closes #3186
- Mockito 4.7.0 -> 4.8.0, supersedes and closes #3185

This commit removes unnecessary cancellation of main source
when fluxTakeUntilOther is terminated by the onComplete of
said main source.

Fixes #3184.

This commit improves `MonoCompletionStage` performance by using handle
instead of whenComplete. The later notably seems to create unnecessary
`CompletionException`.