Performance: chunked MarsInterp, lazy MarsFormat, vinterp/find_n without per-level copies

[mirfjc]

Summary

Performance and memory work on the interpolation path and on MarsFormat. No change in results: every output was checked bit-for-bit (all variables, NaNs included) against the previous code on four files, with the default and with forced small chunk sizes. Builds on #198 (branch aeolis); the three commits here are the ones after it.

Measured on a one-year planetWRF file (7.4 GB, 2676 frames, 60x36x52) on a laptop:

step	before	after
MarsFormat -gcm marswrf	28 s, 20 GB peak	19 s, 2.0 GB peak
MarsInterp -t pstd	53 min, 21 GB peak	76 s, 5.4 GB peak

On a 16-frame file MarsInterp goes from 13 s to 1.5 s (pstd), 6.4 s to 1.3 s (zagl), 9.5 s to 1.5 s (zstd).

What changed

1. vinterp (amescap/FV3_utils.py): Lfull.flatten() and varIN.flatten() (full copies) were called up to six times per output level inside the level loop. They are now taken once as contiguous views before the loop. Same arithmetic, same order of operations.
2. find_n: the index search looped in Python over every column with np.argmin (millions of calls on a year of data). For monotonically increasing input along the first axis, the normal case after reverse_input, the index of the largest X_IN <= X_OUT is now a vectorised count, chunked to bound memory. It gives the same index as the nearest-then-step-down search. Non-monotonic input still takes the original path.
3. MarsInterp time chunking: interpolation is column-local, so the level field, the indices and each variable are computed and written per time slice (Ncdf.log_variable_slice, new). Chunk size is automatic (about 2e7 input elements) or -chunk N. The copy-through of non-interpolated variables is unchanged.
4. MarsFormat lazy processing: the file is opened with dask chunks along time when dask is importable, and the marswrf block was rewritten to stay lazy (no 4D .values), so to_netcdf streams. dask runs with the synchronous scheduler: netCDF4/HDF5 is not thread safe and the threaded scheduler deadlocked on the -ba binning write in the test suite; chunking alone gives the memory benefit. Without dask the behaviour is exactly as before, with a one-line hint. dask is added as an optional extra [large] in pyproject.toml, not a hard dependency. The eta fit for reduced files now uses the first 16 frames instead of all (eta is constant; the 8-frame test file is unchanged).

Verification

• MarsInterp outputs (pstd, zagl, zstd on the full-wrfout bite; pstd on the reduced bite) identical to the pre-change outputs, both with automatic chunking and with -chunk 3 / -chunk 5.
• MarsFormat outputs (default and -stag, both bites) identical to the pre-change outputs.
• pytest tests/test_marsformat.py tests/test_marsinterp.py tests/test_marsfiles.py tests/test_marsvars.py: 60 passed, 4 skipped (with test_marsplot.py).

Not done

• Ames-format (atmos_daily) files were not re-timed here; the same code paths apply, and the find_n fast path triggers whenever the level field is monotonic, which it is for fms_press_calc output.
• Diurn files (time_of_day dimension) go through the same chunked loop (time is still the first axis); tested only via the unit tests.