Fix shape mismatch error in gears caused by varying choose_binset

[OBJECT907]

Issue:

In the gears module, using different binset configurations results in a varying number of tokens. This causes a shape mismatch error when attempting to add positional embeddings, as the sequence length exceeds or differs from the expected dimensions (positional embeddings is always [num_genes, hidden_size])

Fix:

I added a truncation operation to ensure the input sequence length aligns with the model's expectations. The code now explicitly slices the input to keep only the first num_genes tokens before adding positional embeddings.