Add additional warning for the checkpoint loading

[nastiapetrovych]

Previously mentioned in #51, the concern is:

a small warning would still be good to add e.g. "checkpoint has a weight key but this class expects _weight," to avoid a silent trap that goes unnoticed.

[MartinaJoergensen]

Following up with a concrete example of this failure mode: ParT_full.pt's attention weights are stored as one fused in_proj_weight/in_proj_bias per block (standard nn.MultiheadAttention layout), but the PQMultiheadAttention-based model uses separate q_proj/k_proj/v_proj sublayers. load_state_dict(strict=False) had no destination for the fused keys, so they were silently dropped, no warning. Those attention weights trained from random initialization instead of the pretrained checkpoint.

Same pattern as the original report; a checkpoint key with no matching destination, silently ignored. A warning at that point would be genuinely useful for catching this kind of thing early.