Revisit error handling for TaskExtensions.SafeContinueWith

[sharwell]

Version Used: ffa4e6d

Expected Behavior:

1. In production, asynchronous failures should be reported through non-fatal error handling paths

2. During tests, asynchronous failures should result in test failures

   • If the failure occurs while a test is running, the failure should produce a single-test failure (as opposed to a FailFast that crashes the runner)
   • If the failure occurs outside the bounds of a test, it should trigger a FailFast

   I believe the easiest way to implement the above is modifying UseExportProviderAttribute to set FatalError.NonFatalHandler at the beginning of a test. This handler would collect exceptions similar to TestExtensionErrorHandler. At the end of a test, the attribute implementation would reset FatalError.NonFatalHandler to a general handler that triggers a FailFast, and then throw any exceptions that were captured during the current test.

[sharwell]

Non-fatal error telemetry exceptions and stack traces are actually logged (to the servicehub folder) when running inside Visual Studio, which means we are already capturing the desired data in roslyn-integration-CI. Capturing the same details for roslyn-CI involves ensuring something like this is running in unit test cases to log errors:

roslyn/src/VisualStudio/Core/Def/Implementation/Watson/WatsonReporter.cs

Lines 102 to 107 in c8b3af5

	// write the exception to a log file:
	var logMessage = $"[{currentProcess.ProcessName}:{currentProcess.Id}] Unexpected exception: {exception}";
	foreach (var logger in s_loggers)
	{
	logger.TraceEvent(TraceEventType.Error, 1, logMessage);
	}

Once we know the errors are getting logged, we can add the logs as an artifact attached to the build results.