Extend unit tests for metrics to cover resetting the state variables

[jneeven]

Feature motivation

Rather than relying on the default Metric.reset_state() function to correctly reset all variables that should be reset, we should explicitly have a unit test that verifies this for each relevant variable.

Feature description

Since reset_state() doesn't return a tensorflow op, it is not trivial to test this in the current testing ecosystem (calling self.evaluate(metric.reset_state()) will return an error, as reset_state() returns None). Ideally, we should find a way to test this that doesn't compromise ease of use (i.e. we don't want to change the reset function to manually contain a reset operation for each variable, as this doesn't scale well to subclasses with custom variables).