ref(bigquery): Improve BigQuery encoding and logs

[iambriccardo]

This PR removes duplicated and overly verbose logging, and reworks BigQuery row encoding for correctness and efficiency.

Logging

• Downgrades per-step info! logs during pipeline shutdown, replication slot lookup/deletion, and table sync worker startup/state persistence to debug!, since a higher-level summary is already logged elsewhere.
• Removes now-redundant logs that duplicated information already captured by metrics (table copy completion) or by the surrounding log statements.

BigQuery row encoding

• BigQueryTableRow now stores only its already-encoded Protocol Buffer bytes (Vec<u8>) instead of keeping both the source cells and a separate encoded-bytes cache alive. Encoding happens once, up front, in try_from_tagged_cells, and encoded_len/encode_raw just read the already-encoded buffer. Previously, gcp-bigquery-client calling encoded_len() for batch budgeting and then encode_to_vec() (which calls encoded_len() again before encode_raw()) caused every row's date/numeric/UUID/JSON fields to be formatted up to 3 times.
• Removes CellNonOptional/ArrayCellNonOptional from etl::data, which existed solely to give BigQuery a type-level guarantee that arrays contain no NULL elements. BigQuery is now the only destination that cared about this, so the check moved to a runtime validation (validate_array_cell_for_bigquery) that rejects NULL array elements with NullValuesNotSupportedInArrayInDestination, rather than duplicating Cell/ArrayCell in a parallel non-nullable enum.

[coveralls]

coverage: 73.092% (-0.03%) from 73.123% — remove-duplicated-logs into main