feat: add total_samples as a field to logs being emitted

[RobotSail]

Today we emit a number of important values in the logs, but we do not explicitly announce how many samples
we are planning on iterating through. This commit emits total_samples as a field in both the initial log
as well as the subsequent logs during training. This way, this value can be accessed consistently by
any program interested in displaying information about its progress for convenience.

Fixes #325

Signed-off-by: Oleg S [email protected]

[JamesKunstle]

I approved then put the hold- quick q, is len(dataloader.dataset) a cached length value or is it computed each time? we could cache it before we start the loop if it's not cached already.

[RobotSail]

@JamesKunstle What is your concern with the calculation of len(dataloader.dataset) if it's not cached? In Python, the __len__ implementation should be $O(1)$. If this isn't the case then it would be a bug that needs to be fixed elsewhere.

[JamesKunstle]

@RobotSail that tends to be the case for inbuilt stuff but this would be a length calculation for our custom Dataset objects. I confirmed that it's O(1) as best as I can tell so this should be good