Add prepare-stage timing logs to diagnose WindInterpolation bottlenecks

[KULcoder]

Motivation

• Investigate and pinpoint where prepare() spends time for the WindInterpolationModel, with suspicion around file I/O.
• The likely hotspots are input aggregation via xr.open_mfdataset, spline coefficient computation in _splrep/scipy.make_interp_spline (including any dask map_blocks activity), output writes via xr.save_mfdataset, SHA256 hashing of outputs, and metadata dump.

Description

• Instrumented BaseModel.prepare with time.perf_counter() measurements and logger.info messages that report elapsed time per year/month for: open_mfdataset, _prepare_dataset, result.register, and result.dump.
• Added import time and placed timers around the with-block that opens the reference files, the call to _prepare_dataset, the result.register call, and the final result.dump step.
• No behavioral or output format changes were introduced; this is logging-only diagnostic instrumentation intended to reveal whether slowness is I/O- or compute-bound.

Testing

• Ran python -m compileall src/geodata/model/_base.py and the file compiled successfully.

---

Codex Task