fix(#742): match keywords as regex, matching the documented tutorial

[bosd]

Summary

Closes #742.

The template-authoring tutorial has always documented keywords and exclude_keywords as regex patterns — e.g. "Company US and Company\s+US both work". But InvoiceTemplate.matches_input used a plain in substring test, so any keyword that leaned on a regex metachar silently under-matched.

An audit of the bundled templates shows ~16 already write regex-shaped keywords that were effectively broken:

• (?i)Accor, (?i)Digital\sRiver
• BE\s0673\s923
• Express\s+\w+
• Wystawił\(a\)
• NL8152[.]54[.]295[.]B01
• …and more.

Change

• Route both keywords and exclude_keywords checks through a new module-local _keyword_matches helper that calls _regex.search (the compile-once cache used everywhere else in extract/).
• If a keyword doesn't compile as valid regex, fall back to a literal substring check and DEBUG-log the fallback — so a stray unescaped [ or ( in a simple-string keyword like Company [Ltd keeps matching literally instead of breaking.
• Re-export _regex.error (the active engine's error type — re.error by default, regex.error under INVOICE2DATA_REGEX_ENGINE=regex) so the fallback doesn't need to know which engine is active.

Compatibility

• Plain-string keywords with no metacharacters (Acme Corp) behave exactly as before — regex.search(\"Acme Corp\", text) is the same as \"Acme Corp\" in text.
• Keywords that already were regex (bundled templates) now actually work.
• Keywords that were string-but-look-like-regex (rare — Company [Ltd with an unclosed [) still match via the literal fallback.

Test plan

• New unit tests for regex \s+, (?i), alternation, exclude-as-regex, and invalid-regex-falls-back-to-substring.
• Existing test suite passes (only the pre-existing env-related failures remain — test_gvision, test_copy needing pdftotext/tesseract, and QualityHosting.pdf cascade, all red on master too).
• test_default_templates_are_loaded still loads every bundled template (nothing rejected as invalid regex).

https://claude.ai/code/session_01YLnqiMRXGtQQsM7TGRQx6v