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

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