Fix classification labels for batch size 1

[lukasugar]

When using batch_size=1 for smoke tests, squeeze() turns a [1, 1] label tensor into a scalar for batch_size=1,
which breaks CrossEntropyLoss with a batch-size mismatch. view(-1)keeps labels shaped as [B].

For batch sizes other than 1, the behavior remains the same.

[Sllambias]

Seems to also be included in #6 - just porting my comment here:

I am concerned this may cause issues for classification/regression inference. Have you tested that?
(And this shows that the full pipeline tests are long overdue..)

[lukasugar]

Seems to also be included in #6 - just porting my comment here:

I am concerned this may cause issues for classification/regression inference. Have you tested that? (And this shows that the full pipeline tests are long overdue..)

Replied there:

I've tested both classification and regression and it works.
squeeze() and .view(-1) should behave the same - with view have support for scalars - for the current evals/finetuning setup.

Maybe we can close this PR and look at all the changes in the other one