fix(executor): snap entity spans to word boundaries after regex match

[AZERDSQ131]

Closes #7

Implements option (b) from the issue: after substitute_template produces a span's text/start/end, trim leading/trailing whitespace and punctuation and adjust the offsets to match. This turns over-captures like "London," into "London" with correct offsets, so they line up with gold spans under the default text matching mode in evaluation.py::_match_entities instead of counting as both a false positive and a false negative.

Only edge characters are trimmed (whitespace + string.punctuation) — interior characters and the capture group the rule intended are never touched, per the issue's gotcha.

• tests/test_executor.py: 5 new TestSubstituteTemplate cases (trailing punctuation, leading+trailing whitespace, wrapping quotes, no-op when already clean, all-punctuation text left unmodified to avoid producing an empty span) + 1 TestExecuteRegexRule acceptance test matching the issue's example (a rule that over-captures a trailing comma on "London,", asserting the produced span is "London" with offsets that round-trip through the source text).
• Full suite: 317 passed, 1 skipped (pre-existing skip, unrelated).
• ruff check, ruff format --check, mypy clean on the changed files.

[adaamko]

Thanks — the implementation is clean and the offset arithmetic is correct, but I don't think we can merge with string.punctuation as the trim set. It includes %, $, . and /, so correct captures get corrupted: a QUANTITY rule that rightly captures '50%' would be snapped to '50', MONEY '$5,000' to '5,000', 'no. 44774/98' style case references and abbreviations like 'Inc.' also get clipped. Under text matching those all flip from correct to wrong, so the fix for boundary slips would introduce a new class of them.

Suggestion: trim only whitespace plus sentence/wrapper punctuation — something like ,;:!?"'()[]{}<> — and leave $ % . # / & - _ + @ alone ('.' is the debatable one: trimming it fixes 'London.' but breaks 'Inc.'; I'd leave it in the span and let feedback repair handle sentence-final periods). Could you narrow the set, add '50%' / '$5,000' regression tests, and also tick the rights checkbox in the PR body? Happy to merge after that.