CancellableContinuation.invokeOnCancellation cause is inconsistent and breaks its own contract

[qwwdfsad]

Steps to reproduce:

1. This snippet prints null

suspendCancellableCoroutine<Unit> { c ->
    c.invokeOnCancellation {
        println(it) 
    }
    c.cancel()
}

2. Semantically the same snippet prints CancellationException:

suspendCancellableCoroutine<Unit> { c ->
    c.cancel()
    c.invokeOnCancellation {
        println(it) 
    }
}

CompletionHandler contract:

 * The meaning of `cause` that is passed to the handler:
 * * Cause is `null` when the job has completed normally.
 * * Cause is an instance of [CancellationException] when the job was cancelled _normally_.
 *   **It should not be treated as an error**. In particular, it should not be reported to error logs.
 * * Otherwise, the job had _failed_.

[qwwdfsad]

This behaviour seems to be the legacy from the times when CancellableContinuation was an instance of Job (pre 0.22.x versions) and we preserved CompletionHandler ABI.
Tho for CancellableContinuation that does not have invokeOnCompletion the underlying cause is never null.

[qwwdfsad]

One more inconsistency:
1)

suspendCancellableCoroutine<Unit> { c ->
    c.invokeOnCancellation { 
        println("Invoked") // Printed
    }
     c.invokeOnCancellation { expectUnreached() } // Triggers ISE
}

suspendCancellableCoroutine<Unit> { c ->
    c.invokeOnCancellation { 
        println("Invoked")  // Not printed
    }
    c.resumeWith(Result.failure(IllegalStateException()))
}

[elizarov]

As a side note. I've rechecked both in resource-cancel branch (which reworks a significant chunk of CacncellableContinuationImpl machinery):

1. The first inconsistency with cause is still present.

2. The second inconsistency with throwing from invokeOnCancellation is gone, but what should happen if invokeOnCancellation throws an exception is still an open design issue in PR Breaking: Get rid of atomic cancellation and provide a replacement #1937 and that's the reason why it is still a draft.

[dkhalanskyjb]

Both inconsistencies are still there.