Compact parquet per hour partition to preserve external table hive partitioning

[tomaslink]

https://globalfishingwatch.atlassian.net/browse/PIPELINE-4331

---

Summary

Compaction was collapsing the event_hour= partition level when merging a date's parquet files down to one, leaving compacted dates at a different hive-partition depth (event_source=/event_date=) than uncompacted ones (event_source=/event_date=/event_hour=). The pipe-nmea-parsed external table is hive-partitioned over this path, and BigQuery requires every file under the table's URI prefix to match the same partition-key path structure — whether that structure is auto-detected or explicitly declared. Once compaction stopped writing an event_hour= directory for a date, files for that date no longer matched the partition layout the rest of the table used, breaking the external table for that range regardless of partitioning mode.

• Compactor now discovers event_hour= subpartitions per date from GCS and always preserves them when found — compacting each hour independently, output written back under its own event_hour=HH/ folder. This isn't gated by config: it's structurally impossible to accidentally collapse a real hour partition, regardless of how a caller is configured.
• A date with no hour subpartitions at all is compacted as a whole (flat), same as before this fix existed.
• New hourly flag is a narrow assertion, not a mode switch: for event_sources actively expected to be hour-partitioned (e.g. live streaming sources), it asserts a date must have hour subpartitions and raises if none are found — signalling ingestion broke or a stale/wrong declaration, rather than silently falling back to flat compaction. It has no effect when hour subpartitions are actually present — those are always preserved either way.
• Compaction units are modeled explicitly as DailyCompactionUnit / HourlyCompactionUnit (a shared abstract CompactionUnit base, siblings rather than one inheriting the other), moved to their own compact_parquet/units.py module. DailyCompactionUnit.with_hour(hour) derives the corresponding hourly unit for the same date. Compactor.run() builds one flat list of units across the whole date range and loops over it once.
• Bumped gfw-ops version to 0.4.0.

Test plan

• tests/pipelines/compact_parquet/test_main.py covers: hour discovery/sorting, per-unit path()/__str__ for both unit types, with_hour() derivation, the preserve-hours-regardless-of-flag behavior, the hourly=True-with-no-hours-found hard failure, and the existing swap/manifest/retry/copy-mode flows against the new unit-based API
• Full gfw-ops suite passes (101 tests)
• ruff/mypy show only pre-existing, unrelated debt (mypy actually dropped from 9 → 7 pre-existing findings as a side effect of this refactor) — no new violations introduced.