Skip metric computation if output is None

🚀 Feature

Considering the context of https://github.com/pytorch/ignite/pull/1024
and another point about using tensor.item() on each iteration with TPU which slows down the computation, we need to have a way to skip using state.output if set to None.

In training on TPU

def training_step(e, b):
    # ....
    
    # instead of returning each iteration loss value
    # return loss.item()
    # we would like to do it 
    return loss.item() if (e.state.iteration - 1) % 20 == 0 else None 

however, if we have a running metric attached to the output it will fail if encounter None.

So, idea is to check engine.state.output here and return before doing any metric's update:
https://github.com/pytorch/ignite/blob/e3fc04e147db4c97a2ceb70597366d562e15c96f/ignite/metrics/metric.py#L123-L124

@sdesrozis @erip any thoughts ?

This isssue is related to https://github.com/pytorch/ignite/issues/996 but the difference is that we still want to execute all other handlers on iteration even if output is None.

Start in ignite/metrics/metric.py at the state.output handling referenced by the issue, and trace how metric updates are invoked during an iteration. Verify the behavior with an output of None: metric updates should be skipped while other iteration handlers still run; check the surrounding metric tests for the appropriate coverage.