Implement FeatureExtractor.expression() for email domain/TLD extractors #145

Context

Raised by @tim-band during review of #137 (discussion): EmailDomainExtractor.expression() and EmailTopLevelDomainExtractor.expression() currently just raise NotImplementedError(). Tim sketched the actual per-dialect SQL:

Postgres: SELECT SPLIT_PART(email, '@', 2) AS domain FROM customer limit 20
MsSQL: SELECT SUBSTRING(email, CHARINDEX('@', email) + 1, LEN(email)) AS domain FROM customer WHERE CHARINDEX('@', email) > 1

and separately floated a more general AfterCharacter dialects.py function class that could serve both domain and TLD extraction (TLD being "after the last . of whatever's after @").

Why this doesn't bite today

EmailDomainExtractor/EmailTopLevelDomainExtractor are only ever used with CategoryBuilder (see EMAIL_PIPELINE in statistical_fidelity.py), and CategoryBuilder.build_from_table doesn't currently call .expression() at all - it samples rows and calls .extract() in Python. So the NotImplementedError is unreachable in the current architecture.

Why it's still worth doing

It becomes a real gap the moment #144 (moving CategoryBuilder onto SQL-side GROUP BY for extractors that support it) is implemented - at that point, the EMAIL profile's "domain"/"tld" pipeline stages would need these to work, or would have to keep a Python-sampling fallback specifically for them.

Proposal

Add dialect-aware SQL expression(s) to dialects.py (Postgres SPLIT_PART/MSSQL SUBSTRING+CHARINDEX for the "after @" case, reused for "after the last ." for TLD), matching the existing pattern of SecondsDifference/WordCount/SentenceCount. Consider Tim's generalized AfterCharacter(column, char) form if it cleanly covers both call sites.

Depends on / pairs with #144.