fix(finetune): fix variable shadowing that broke per-predictor trace filtering

[isaacbmiller]

The loop variable pred_ind in _prepare_finetune_data shadows the function
parameter of the same name. This causes the filter condition
pred_ind == pred_ind to always evaluate to True (self-comparison),
meaning all predictor traces are included regardless of the pred_ind
parameter.

When multitask=False, each predictor should only receive its own traces
for fine-tuning, but due to this shadowing bug, every predictor receives
all traces — effectively behaving the same as multitask=True.

Fix: rename the loop variable to trace_ind to avoid shadowing.