Detect mismatched CF scale/offset on concat via codecs

[TomNicholas]

What

Closes #1004.

Concatenating virtual datasets whose source files were CF-packed with different scale_factor/add_offset previously succeeded silently, keeping only the first array's attributes and corrupting decoded values for every chunk sourced from a non-first file.

This fixes it entirely within the HDF parser by representing CF scale/offset packing as the new zarr scale_offset + cast_value codecs (from zarr-developers/zarr-python#3874) and stripping the corresponding attributes. Because the packing now lives in the codec pipeline, files packed with different parameters produce different codec pipelines, and the existing check_same_codecs check refuses to concatenate them. No concatenation code was changed.

How

• parsers/hdf/filters.py — cf_codecs_from_dataset builds ScaleOffset(scale=1/scale_factor, offset=add_offset) + CastValue(<int dtype>) and reports the decoded float dtype.
• parsers/hdf/hdf.py — wires the codecs into the array metadata (before the byte-level filters), decodes the fill value into the float domain, and drops scale_factor/add_offset from attrs.
• writers/kerchunk.py — kerchunk targets zarr v2/numcodecs, which can't represent these codecs, so the writer recognises them and reverts to scale_factor/add_offset attributes + the integer dtype on write, where xarray's decode_cf reapplies them on read.
• pyproject.toml — adds cast-value-rs to the hdf extra (required by cast_value at parse time).

Tests

• TestConcatMismatchedCFEncoding (HDF): mismatched scale_factor/add_offset raise on concat (surfaced as the ScaleOffset codec mismatch from check_same_codecs); matching encoding round-trips through icechunk with correct decoded values.
• Full suite green.

Known caveat (feedback wanted)

Codec-based decoding computes packed / scale, whereas xarray's decode_cf computes packed * scale_factor. These agree to within ~1 ULP but are not bit-identical for non-power-of-two scale factors (the scale_offset codec only offers a divide on decode). This is below the quantization floor that packing already imposes, but it means two loadable-variable tests now compare CF-packed variables with assert_allclose rather than assert_identical. See related discussion in zarr-developers/zarr-python#3874.

TODO before un-drafting

• Decide whether the ~1 ULP decode_cf divergence is acceptable as-is or warrants raising upstream on zarr-python#3874.
• Regenerate pixi.lock for the new cast-value-rs dep.
• Consider whether DMR++ / other parsers should share the CF→codec path.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.94%. Comparing base (496fd0f) to head (09d374b).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1010      +/-   ##
==========================================
- Coverage   89.97%   89.94%   -0.03%     
==========================================
  Files          36       36              
  Lines        2224     2278      +54     
==========================================
+ Hits         2001     2049      +48     
- Misses        223      229       +6     

Files with missing lines	Coverage Δ
virtualizarr/parsers/hdf/filters.py	98.43% <100.00%> (+0.45%)	⬆️
virtualizarr/parsers/hdf/hdf.py	95.80% <100.00%> (+0.21%)	⬆️
virtualizarr/writers/kerchunk.py	77.27% <100.00%> (+4.44%)	⬆️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.