v0.8.9 — Opt-in Indic PII masking (Verhoeff + Devanagari)

Added — Opt-in Indic PII masking: Verhoeff + Devanagari (v0.8.9)

A new pii_locales opt-in tag on sanitize_output() / mask_sensitive_data()
/ detect_sensitive_data() / AirlockConfig. Defaults to [] so the
zero-dep, US-default PII surface is unchanged for existing callers. When
pii_locales=["in"] is supplied, three things activate:

1. Aadhaar Verhoeff checksum gate. The existing Aadhaar regex
   (\b[2-9][0-9]{3}[\s-]?[0-9]{4}[\s-]?[0-9]{4}\b) is permissive —
   any 12-digit number starting 2-9 matches. With the opt-in, each
   match is validated against the UIDAI Verhoeff checksum (the same
   algorithm UIDAI uses to compute the last digit). Cuts the FP rate
   ~10x on random 12-digit IDs / phone numbers.
2. Devanagari personal-name detection — new
   SensitiveDataType.PERSONAL_NAME_DEVANAGARI enum member. Regex
   matches sequences of 2+ Devanagari word characters (Unicode block
   U+0900–U+097F), optionally joined by whitespace. A small
   common-noun allowlist (greetings, pronouns, interrogatives, basic
   connectives, copulas) filters obvious false positives — multi-word
   spans where every token is in the allowlist are dropped. This is a
   conservative heuristic, not NER; production callers who need precise
   name extraction should layer a proper NER on top.
3. India PII auto-include in sanitize_output(mask_pii=True):
   Aadhaar / PAN / UPI / IFSC / Devanagari are added to the default
   masked types. Without the locale tag, callers must still pass
   types=[...] explicitly to mask India-locale PII — that
   behavior is unchanged.

The existing Aadhaar / PAN / UPI / IFSC enum members, regex patterns,
default mask strategies, and tests are untouched — the new gate
is purely additive. Reuses the existing four masking strategies
(FULL / PARTIAL / TYPE_ONLY / HASH); no new strategy.

New public exports: SensitiveDataType.PERSONAL_NAME_DEVANAGARI,
AirlockConfig.pii_locales field. The new pii_locales parameter on
detect_sensitive_data / mask_sensitive_data / sanitize_output is
keyword-only by convention (positional callers from before are unaffected).

pii_locales is also accepted in TOML config (pii_locales = ["in"]
or pii_locales = "in,us").

32 new tests in tests/test_indic_pii_masking.py cover: Verhoeff
helper (valid / invalid / non-digit / mutation), Aadhaar locale gate
(opt-in passes valid + drops invalid; no-opt-in keeps permissive
behavior + separator normalization), Devanagari detection (single +
multi-word names; common-noun allowlist FP filtering; mixed
greeting+name spans), sanitize_output end-to-end with the locale
extending the default PII set, full @Airlock integration via
AirlockConfig.pii_locales, and backwards-compatibility (all old
call shapes still work).