Don't send out Runtime error telemetry when can't create LearningModelDevice on machine without hardware adapters

[ryanlai2]

Runtime error telemetry should only be sent out for scenarios where WinML behaved in an unexpected incorrect way.

In the scenario, that there are no hardware adapters available and a LearningModelDevice targetting GPU can't be created, Runtime error telemetry shouldn't be sent but hresult error should still be returned to the client app.

TelemetryFallback callback function that calls LogRuntimeError for any failure can be removed for these reasons:

• WinML "THROW_IF_FAILED" macros already call LogRuntimeError, no need to send error telemetry twice.
• The telemetry data sent by this callback function doesn't contain line and file information.
• Having this prevents WinML from throwing error without sending telemetry.