Refactor BLR transfer to remember batch effects

[contsili]

Related to #382

1. Sex is a global covariate, consistent across datasets → transferable.
   Sites are dataset-specific; every new dataset introduces new sites → not transferable.

Action: In self.be_idx_gen keep the lists that you want to transfer to (eg sex). We could also allow the user to decide here

2. What to do when some sites appear in both train and target datasets?

Idea: At transfer time, compare the source scanner list against the target dataset's scanner list
Known scanners (overlap): reuse learned parameters directly
New scanners: re-initialize as new BE

3. What to do when the transfer dataset has unknowns?

Idea: keep the learned parameters from the training dataset?

4. What we do when the transfer dataset has more or less BE than the train dataset?

Ideas: When the transfer dataset has less give a warning. For more proceed normally

Test code:
Try splitting the fcon data see https://pcntoolkit.readthedocs.io/en/latest/pages/tutorials/06_transfer_extend.html

If i want a big dataset i can use the lifespan data