Optional: write zarr alongside netCDF for direct-from-S3 reads

[joglekara]

Not needed for v1 of the Thomson analysis browser — filed so the option stays visible when weighing how much server we want to run long-term.

The browser's backend will download .nc artifacts and slice them server-side. If fits also wrote the same datasets as consolidated zarr, a web client could read chunks straight from S3 (zarr-js / zarrita) and most server-side slicing endpoints become optional — the visualizer trends toward a static frontend plus the MLflow REST API.

Scope

• Config flag (default off) to write to_zarr copies with consolidated metadata next to the existing netCDF in binary/
• Chunking chosen for the access patterns: whole-lineout reads and downsampled overview reads
• Entry in the artifact manifest (see the manifest/contract issue)
• Check artifact-size impact on public-ergodic-continuum before ever defaulting it on

netCDF stays canonical either way.