[GSoC] Integrate the trained model to inject required phrases into rules

[Kaushik-Kumar-CEG]

follows #5137. uses the trained model from there

Goal

Run the finetuned model over eligible rules that have no required phrases, and
inject the predicted {{ }} markers using scancode's existing helpers

Tasks

• load the trained model (PyTorch, fp32) and run inference on rule text
• pick target rules with get_updatable_rules_by_expression (skips ineligible ones)
• rebuild predicted phrases from the BIOES spans
• validate every phrase with is_good() before writing
• inject with add_required_phrase_to_rule, tagged source=ml_model
• --dry-run and a confidence threshold so only confident phrases auto-inject
• note to run scancode-reindex-licenses after any write