Don't report cancelled remote transfers as errors

[fmeum]

Description

When a bulk transfer to or from the remote cache is cancelled, every one of its in-flight transfers fails with a CancellationException. mergeBulkTransfer subscribes to all transfers of a bulk operation at once via flatMapSingle, which terminates on the first error and hands every subsequent one to RxJava's global error handler.

RemoteModule installs a global error handler that reports those as error events, so a single cancelled bulk transfer printed a java.util.concurrent.CancellationException: Task was cancelled stack trace for every transfer but the first.

A transfer is only ever cancelled as part of tearing down the operation it belongs to, so toTransferResult now maps CancellationException to the same result as InterruptedException instead of propagating it downstream.

This PR also replaces two isDisposed()-then-onError() sequences in RxFutures and AsyncTaskCache with tryOnError. Those checks aren't atomic with the delivery: the emitter can be disposed in between, in which case onError hands the error to the global error handler as well.

Motivation

Users see spurious ERROR: java.util.concurrent.CancellationException: Task was cancelled stack traces, one per concurrently cancelled transfer, whenever a build that is uploading or downloading blobs is interrupted or an action fails while a bulk transfer is in flight.

This also shows up as a flake in RemoteExecutionServiceTest.uploadInputsIfNotPresent_interrupted_requestCancelled, which fails whenever RxNoGlobalErrorsRule observes the leaked exception. That test failed in roughly 2-5% of runs before this change and passed 100 out of 100 times after it.

Build API Changes

No

Checklist

• I have added tests for the new use cases (if any).
• I have updated the documentation (if applicable).

The existing uploadInputsIfNotPresent_interrupted_requestCancelled covers this, but only probabilistically: the number of transfers that are cancelled concurrently determines whether an error is leaked, and the tryOnError changes fix races that can't be triggered deterministically.

Release Notes

RELNOTES: None

[fmeum]

@bazel-io fork 9.3.0