Account for size of token replacements when sizing the bar

[FremyCompany]

If we are not doing this, then the bar can overflow to the next line, then it starts drawing itself over and over again, because the start of the line is not the real start of the line but the start of the wrapped line, and that is really annoying...

Since we want the bar size to remain stable, the maximum size for a token is remembered.

[FremyCompany]

This was prompted by the line wrapping issue happening in tensorflow.js and driving me nuts.

cc @bileschi @caisq (if this isn't merged, the fix is to make the name of the token dependent of its expected size, which sucks but sucks less than outputting a new progressbar for every batch)

[caisq]

@FremyCompany Can you file an issue at TensorFlow.js' Issues page with details on the model you are using?

[FremyCompany]

@caisq Sure. This would happen for any model that has a few outputs/metrics.

[caisq]

@FremyCompany Got it. I think I know what the problem is, but will appreciate you filing an issue formally regardless. Thanks.