Unify input file formats to Zarr and generalize dataloaders

[grey-nearing]

Summary of Changes

This PR standardizes input file storage and data loaders across the codebase onto the Zarr format, improves data loader flexibility, and provides seamless migration tooling:

1. Unified Zarr Input Format & Generalized Dataloaders:
   • Standardized Caravan static attributes (attributes.zarr) and streamflow time series (streamflow.zarr) into consolidated, chunked Zarr stores.
   • Generalized the Multimet dataloader with robust path resolution—supporting single consolidated stores (timeseries.zarr), per-product subdirectories (<product>/timeseries.zarr), and case-insensitive store matching.
   • Removed hardcoded basin naming assumptions (such as basin.split('_')[0]), allowing arbitrary basin identifier naming conventions.
2. Feature Scaler Migration (scaler.zarr):
   • Migrated feature normalization scalers to Zarr format (scaler.zarr) while retaining backward-compatible fallback to existing scaler.nc files.
3. Caravan Dataset Conversion Utilities:
   • Introduced googlehydrology.datautils.convert (convert_caravan_attributes, convert_caravan_timeseries, convert_caravan_to_zarr) for converting legacy NetCDF and CSV directories into consolidated Zarr stores.
   • Exposed the conversion tool via the CLI entrypoint: run convert-caravan.
4. Output Zarr Store Metadata Consolidation:
   • Updated BaseTester.evaluate() to automatically write consolidated metadata for {period}_results.zarr at the completion of inference.
5. CMAL Sampling & Negative Sample Handling:
   • Extended sample_cmal_deterministic to accept scaler and apply analytical quantile zero-bounding through _handle_negative_values.

---

Verification & Testing

• Numerical Parity: Validated identical inference results between legacy NetCDF inputs and converted Zarr stores on both the 5-basin tutorial model and the full 615-catchment FloodHub operational benchmark.
• Test Suite: All 212 unit and end-to-end integration tests pass successfully (pytest test/).