Refactor EpochMetric and make it idempotent

[sadra-barikbin]

By adding an assertion to check recomputing metric does not change its result, _test_distrib_integration fails.

[vfdev-5]

Seems like there is a bug here:

ignite/ignite/metrics/epoch_metric.py

Lines 148 to 152 in 7c8d801

	if ws > 1 and not self._is_reduced:
	# All gather across all processes
	_prediction_tensor = cast(torch.Tensor, idist.all_gather(_prediction_tensor))
	_target_tensor = cast(torch.Tensor, idist.all_gather(_target_tensor))
	self._is_reduced = True

_prediction_tensor and _target_tensor are temporary tensors but self._is_reduced is set to True as for other metrics where @sync_all_reduce decorator updates metric's attributes and thus we avoid duplicated all_reduce calls.

Thanks for catching that @sadra-barikbin !

[vfdev-5]

LGTM, thanks @sadra-barikbin