Tolerate null created/modified in searchable extract

[jp-spotter]

Problem

The searchable metadata extract crashes mid-run (KeyError: 'created') when the ThoughtSpot API
omits a timestamp for certain object types — observed: FEEDBACK dependents omit created,
LOCAL_GROUP groups omit modified. One missing field aborts the whole extract, so affected
clusters get no data at all.

Fix

Treat created/modified as a nullable class in the searchable models, and coerce through a
shared helper that tolerates absent/null:

• validators.utc_from_millis() — epoch-millis → UTC datetime, or None when absent (guards the
  /1000 against KeyError/TypeError).
• validators.ensure_datetime_is_utc now short-circuits None, so model validators inherit it.
• created/modified → Optional on User, Group, Tag, DataSource, MetadataObject,
  DependentObject; transformers use src.get(...) via the helper.

A batch extract should capture whatever is available (null included) rather than crash.

No contract change

• DB columns were already nullable (SQLAlchemy default) — no schema migration needed.
• Genuinely-required timestamps in other tools still reject None loudly (their fields stay
  non-optional; pydantic rejects at the type check).
• Return shape unchanged.

Validation

• New unit tests (tests/test_searchable_transformer.py); full suite green.
• Audited the live transformers against two clusters (~140k records) — 0 failures, including
  the LOCAL_GROUP/modified case that previously failed.

Scope

Searchable only. The shared validators change enables the same treatment for the other tools'
transformers (archiver, scriptability, user-management, …) as a follow-up.