Add script to predict and mark required phrases in license rules

[Kaushik-Kumar-CEG]

part of #5243. Adds the tool that runs the trained tagger from #5222 over license
rules and marks the required phrases it predicts with {{ }}

74% of license expressions have no required phrase coverage, this fills them in

this PR:

• add_ml_phrases.py - picks the rules that have no phrases, predicts the spans
  and marks them. --model (local dir or hf id), --license-expression,
  --dry-run, --limit, --verbose. fp32, cpu or gpu
• test_add_ml_phrases.py - unit tests, the tagger and tokenizer are faked so
  they need no weights and no network

rule loading, validation and marking all go through the existing helpers in
required_phrases.py, and the words come from required_phrase_splitter so they
line up with what the model was trained on. nothing in src/ changes

needs #5222 for PhraseTagger and the label scheme. it never reindexes, it just
prints a reminder once it has written something

notes

rules are loaded once instead of calling get_updatable_rules_by_expression per
expression, that reparses all 36k rule files on each call and returns nothing when
passed None. i can fix both there in a separate PR if you prefer

rules that already have {{ }} are skipped, 9875 of them. the braces are tokens
too so they shift every word after them and the tags stop lining up

checks

--dry-run over apache-2.0, mit, gpl-2.0 and bsd-new leaves every rule file
unchanged. a real write reloads fine, keeps the ignorables, and the rule is
skipped on a second run so an interrupted run resumes. feeding the gold labels
back in as a perfect prediction, every phrase is still found by
find_phrase_spans_in_text and 97% pass is_good

the checkpoint is not published yet, so the download path in load_model has not
been run against real weights

disclosure : used Claude to help review and clean up bugs in script