Use Duration as the ground truth for communicating durations

[dkhalanskyjb]

Historically, the library evolved using "a Long of milliseconds" as the standard of denoting durations.
Since then, kotlin.time.Duration appeared, encompassing a number of useful conversions.

There are several consequences to this change.

• Before, delay(Long) and delay(Duration) were not easily expressed via one another. For example, delay(Long.MAX_VALUE / 2 + 1) (up until Long.MAX_VALUE) used to be considered a valid delay, but it was not expressible in delay(Duration). Therefore, delay(Long) was the more fundamental implementation. However, delay(Duration) could not just be expressed as delay(inWholeMilliseconds), as we need to round the durations up when delaying events, and this required complex logic. With this change, delay(Duration) is taken as the standard implementation, and delay(Long) is just delay(timeMillis.milliseconds), simplifying the conceptual space.
• The same goes for other APIs accepting either a duration or some Long number of milliseconds.
• In several platform APIs, we are actually able to pass nanoseconds as the duration to wait for. We can now accurately do that. This precision is unlikely to be important in practice, but it is still nice that we are not losing any information in transit.
• On Android's main thread, it's no longer possible to wait for Long.MAX_VALUE / 2 milliseconds: it's considered an infinite duration. Long.MAX_VALUE / 2 - 1 is still fine.
• In kotlinx-coroutines-test, before, it was possible to observe correct behavior for up to Long.MAX_VALUE milliseconds. Now, this value is drastically reduced, to be able to test the nanosecond precision.
• In kotlinx-coroutines-test, we now fail with an IllegalStateException if we enter the representable ceiling of time during the test. Before, we used to continue the test execution, only using the order in which tasks arrived but not their virtual time values.

[dkhalanskyjb]

It's a draft to signify that I'm not sure this change is needed and want a second opinion. The simplified code looks nicer to me, but other than that, it's technically a breaking change without a clear user need. I mostly did this to thoroughly familiarize myself with time-based operations in our library.

[qwwdfsad]

I am generally okay with the approach (probably we'd like to target a major release here).

Maybe we should also align it with all-compatibility and provide a migration window just in case.

Note that I'd rather skimmed through than reviewed carefully, will do when you'll promote it to the PR