Intruder detection scoring and example embedding scoring

[SrGonao]

This PR became bigger than I wanted because by mistake I added code from two different branches.

It adds:

• Two new scoring methods, that don't require any model explanations.
• Centers activating examples. For now, 3/4 of the window will be to the left of the maximal activating token in the previous windows, and 1/4 will be to the right. To do this, I'm discarding all activations from the first and last window of each batch (which is 1/4 of the total activations we have if we collect with ctx_len 256).
• Some fixes to the online client
• New example sampling option