Add NIOAsyncWriterSinkDelegate._didSuspend hook for testing

[simonjbeaumont]

Motivation:

The tests for the NIOAsyncWriter currently rely on a number of Task.sleep calls to wait for tasks to be suspended or cancelled before performing some other operation. These have been shown to be flakey in CI. We could just increase all the sleeps, but instead we can make use of XCTestExpectation to wait for things to happen. The events that we are waiting on are these:

1. Waiting within a task for the task to be cancelled before we yield.
2. Waiting outside a task that is performing a yield for that task to be suspended.

This didn't require any additional changes to the writer for (1) and can just wait for an expectation inside the task that we fulfil after the cancelling the task from outside.

For (2) we need to have a hook that we can use when the task has performed the state transition and is parked that we can use to fulfil the expectation.

Modifications:

• Add an internal _didSuspend optional closure to NIOAsyncWriterSinkDelegate.
• Replace all calls to Task.sleep(nanoseconds:) in tests with await fulfillment(of:[timeout:]).

Result:

Tests should be more deterministic.

[FranzBusch]

I hate it but it solves the flakiness, so let's do it! We should probably do the same for the NIO[Throwing]AsyncSequenceProducer tests. Do you mind doing the same there? Can even put it into this PR if you want!

[FranzBusch]

I prefer if we don't use @testable. It is a horrible compiler hack and I would love to slowly get away from it by either using SPI or the new package modifier. For now let's use @_spi(Testing) and we can migrate that in the future.

[simonjbeaumont]

OK, discussed this offline.

I think we agreed that what we really want here is package, but we cannot use that today. Making it SPI might have a performance impact when compiled without -enable-testability because we'd have to make some things public (SPI) that were internal, which could impact the optimisations possible.

I'm pretty flexible at this point. I have a lean toward @testable import because it leaves the release build unaffected, but if you'd like I can do the SPI dance before we merge.

[simonjbeaumont]

I hate it but it solves the flakiness, so let's do it! We should probably do the same for the NIO[Throwing]AsyncSequenceProducer tests. Do you mind doing the same there? Can even put it into this PR if you want!

Added commit to give the NIOThrowingAsyncSequenceProducerTests the same treatment.