Fix validation loss aggregation in VSUNet

[ziw-liu]

Fix a bug where the training loop does not clear loss/validate metrics after logging it. The logged metric will be the average of all the epochs, affecting checkpoint selection.

[ziw-liu]

This bug means that previously, as long as the loss of the new $N$-th epoch $l_N$ is lower than the average loss than the previous epochs, the new 'validation loss' logged (which is actually the average over $N$ epochs, $\overline{l}_{N}$), will appear to decrease:

$$\overline{l}_{N} - \overline{l}_{N-1} = \frac{\sum_{i=0}^N l_i}{N} - \frac{\sum_{i=0}^{N-1} l_i}{N-1} = \frac{1}{N} (l_N - \frac{1}{N-1} \sum_{i=0}^{N-1} l_i) = \frac{1}{N} (l_N - \overline{l}_{N-1})$$

[ziw-liu]

These being said, we also watch per-step validation losses. They are a bit noisy, but I don't think our models have major overfitting issues.

[edyoshikun]

more users more bugs! thanks for the fix.