Fix last_hidden_state slicing position in ModernBertForSequenceClassification

[whitebox2]

Problem

The shape of the tensor passed to the classifier is different.
Transformers: correctly uses the [CLS] token → shape (batch_size, hidden_size)
slices across the feature dimension → shape (batch_size, seq_len)

Suggestion

I would like to propose the following change.
before

let last_hidden_state = output.i((.., .., 0))?;

after

let last_hidden_state = output.i((.., 0, ..))?;

This selects the first token (i.e., [CLS]) across all samples , resulting in a shape of (batch_size, hidden_size)

issue

#3000