Add method to join new CMIP6 data into existing parquet data catalogs

## Summary

Currently, parquet data catalogs (e.g. `cmip6_catalog.parquet`) are regenerated from scratch by scanning entire directory trees via `generate_esgf_catalog.py` or `generate_catalog()` in `solve_helpers.py`. There is no way to incrementally add new CMIP6 datasets to an existing parquet catalog without re-scanning everything.

As ESGF archives grow, full regeneration becomes increasingly expensive. We need a method to join/append new data into an existing parquet catalog.

## Current Behavior

- `generate_catalog()` scans all directories and produces a complete DataFrame
- `write_catalog_parquet()` overwrites the parquet file
- `load_solve_catalog()` reads the parquet file into a DataFrame
- No incremental update path exists — adding one new dataset requires re-scanning the entire archive

## Proposed Behavior

Add a method (e.g. on `DatasetAdapter` or as a standalone helper in `solve_helpers.py`) that:

1. Loads the existing parquet catalog
2. Scans only the new directory/files for metadata (using the existing `find_local_datasets` machinery)
3. Merges the new entries into the existing catalog, deduplicating on a suitable key (e.g. `instance_id` or the combination of DRS facets)
4. Applies version filtering (`filter_latest_versions`) so superseded versions are dropped
5. Writes the updated catalog back to parquet

### API sketch

```python
def update_catalog_parquet(
existing_path: Path,
new_directories: list[Path],
source_type: SourceDatasetType,
strip_path_prefix: Path | None = None,
) -> pd.DataFrame:
"""Join new datasets into an existing parquet catalog."""
...
```

This could also be exposed as a CLI option, e.g.:
```bash
python scripts/generate_esgf_catalog.py --update --cmip6-dir /path/to/new/data
```

## Considerations

- Deduplication key: need to decide what constitutes a "same dataset" for merge purposes (likely the full set of DRS facets minus version, so newer versions replace older ones)
- Should handle the case where the existing parquet doesn't exist yet (fall back to full generation)
- Version filtering should run after the merge so that new versions of existing datasets correctly supersede old ones
- The `{data_dir}` path prefix stripping should be consistent between old and new entries