Ref #94

What this does

• select_forcing_backend() auto-switches netcdf/zarr based on domain size and timestep count (small domain, many timesteps → zarr)
• zarr_data_extract() uses Dask for lazy, chunked reads from a zarr store
• Weighting logic is shared between both backends, no duplicated code
• build_combined_zarr_reference() combines individual kerchunk references into one dataset when given a list of URLs
  Note
  Pins zarr<3 — zarr≥3 breaks reading kerchunk reference filesystems (confirmed on 3.2.1). kerchunk needs zarr≥3 to import, so it's imported lazily inside build_combined_zarr_reference() rather than as a top-level dependency, to avoid the conflict.

Not in this PR

• Oldest-lookback selection logic for operational use
• Auto-building the zarr store from a raw bucket (vs. a list of URLs)
• Action item Revert "Forcingprocessor workflow" #2 from Zarr in Forcing Processor #94 (AORC / generalized forcing_url tool) — separate scope