feat(verify): gate new-device candidates before device_catalog

[Seungpyo1007]

Adds python -m app.verify candidates. It is a staged accept/hold gate for new-device rows, here the Google Play hold queue from ADR-016, applied before they become device_catalog entries.

Stages:

• junk
  • vendor list, non-mobile Console form factor, TV/panel SoCs and 4K screens, TV/box/POS names and codenames;
  • vendor mix: a brand whose Console rows are mostly non-mobile can't vouch for its rows that have no Console data.
• seller
  • ODM/distributor list;
  • brand-level share of rows that name another brand;
  • a row naming a different brand is held.
  • A row naming exactly one sibling brand is refiled (TCL under alcatel).
• label: market suffix (the clean name is taken from models when present), codename echo, non-Latin script, bare model codes for major brands, internal strings.
• dup
  • exact id across all brands, including variant.model_numbers;
  • region-normalised id;
  • name tokens inside one of the brand's slugs;
  • major brands only: numeric stem, and rows with sdk ≤ 25, which map to pre-id imports.
• batch: duplicate model numbers within the input.

Nothing is dropped; held rows keep their reasons for review. Unit tests are in tests/verify/test_catalog_candidates.py.

Refs #98