Add whole-run zarr outputs across inference segments

[elynnwu]

Segmented inference writes each segment's outputs to its own directory, which fragments long-run outputs (e.g. monthly means) into per-segment files. This PR makes zarr outputs configured through DataWriterConfig.files accumulate into a single whole-run store shared by all segments, written to any fsspec-compatible location (including directly to a remote bucket), so a preempted-and-resumed run produces the same single stores as an uninterrupted one. netCDF outputs are unaffected and MonthlyDataWriter / save_monthly_files are untouched.

Block-mean time_coarsen needs no special handling across segments: it is stateless per batch and its validation requires the per-segment step count to be divisible by the coarsen factor, so no coarsening block can straddle a segment boundary; whole-run store offsets are simply expressed in coarsened units.

Changes:

• fme.ace.inference.data_writer.zarr.ZarrWriterAdapter and SeparateICZarrWriterAdapter take a start_timestep: the store is sized for the whole run, each segment region-writes its slice of the time axis, and resumed segments open the existing store in append mode; only the run's first segment may create the store, since its initial condition times determine the store's time coordinates

• fme.ace.inference.data_writer.monthly_zarr.MonthlyZarrWriter, built from a files entry with zarr format and monthly_mean time coarsening (previously rejected), with the month axis pre-allocated from the whole run's length; because monthly accumulation is not idempotent, finalize writes a netCDF snapshot to the segment directory (before the restart files, so restarts remain the completion marker) and the next segment's writer restores from it, making segment re-runs exact

• fme.ace.inference.data_writer.zarr.ZarrWriterConfig gains an optional path for placing stores outside the experiment directory

• fme.ace.inference.data_writer.segment.SegmentContext threaded at runtime from the coupled segmented driver through DataWriterConfig / FileWriterConfig build paths; no user-facing config changes and non-segmented runs are unchanged

• Tests added

[elynnwu]

The idea for this is that we can write the monthly zarr to gcs like this:

data_writer:
  ocean:
    save_prediction_files: false
    save_monthly_files: false
    files:
      - label: output               # -> gs://bucket/output.zarr
        save_reference: false       # predictions only; no "_predictions" suffix
        time_coarsen:
          method: monthly_mean
        format:
          name: zarr
          path: gs://bucket         # store dir; {label}.zarr is created inside
        # names: [sst, o_prog]      # optional: subset of variables to save
  atmosphere:
    save_prediction_files: false
    save_monthly_files: false
    files:
      - label: output               # -> gs://bucket/atmosphere/output.zarr
        save_reference: false
        time_coarsen:
          method: monthly_mean
        format:
          name: zarr
          path: gs://bucket/atmosphere