fix(mipro): store demos_idx instead of instruction_idx in raw_chosen_params

[isaacbmiller]

The raw_chosen_params dict incorrectly stores instruction_idx for the
{i}_predictor_demos key instead of demos_idx. This is a copy-paste error

• the line immediately above correctly stores demos_idx in trial_logs.

This causes Optuna's Bayesian model to learn wrong associations between demo
choices and scores during minibatch optimization, silently degrading
optimization quality.