Propagate drop_variables into HDFParser subgroup recursion

[aladinor]

Closes #1057

Problem

HDFParser(drop_variables=[...]) filtered variables only at the group the parser was called on. _construct_manifest_group's recursive call for child groups omitted the parameter, so it reset to None at every level of descent — a dropped name nested one group deeper was still parsed. If that nested dataset has a dtype zarr cannot represent (e.g. a compound dtype with a vlen-string field), the whole parse failed even though the user explicitly excluded it. Real-world case: ODIM_H5 weather-radar volumes (FMI's open archive, files from 2026 on) carry compound-dtype quality*/legend tables under every sweep group, making any root-group parse impossible.

Fix

The user-supplied drop_variables set now propagates into the subgroup recursion, so a listed name is dropped wherever it appears in the walked hierarchy — the semantics the parameter's docstring already implied, and the same behaviour as the Zarr/Icechunk parsers' recursive walks (construct_manifest_group_tree applies skip_variables at every level).

One subtlety: the per-group auto-detected non_coordinate_dimension_vars (the big-endian workaround set) are deliberately not propagated — they are level-specific, and leaking them into sibling subtrees could silently drop an identically named, perfectly valid variable. They are now applied only to the current group's datasets comprehension, where they could previously never match a child group's name anyway (they are always dataset names, and HDF5 group namespaces are unique).

Tests

• test_drop_variables_in_nested_groups — the regression: parsing the nested compound-dtype file fails without the drop, succeeds with it (new nested_compound_dtype_hdf5_url fixture mirroring the issue MRE)
• test_drop_variables_applies_at_every_depth — a name present at two depths is dropped at both
• test_drop_variables_group_name_skips_subtree — a group name in drop_variables skips the whole subtree
• test_dimension_var_drops_stay_group_local — an auto-dropped dimension var in one group does not drop an identically named variable in a sibling group

manifest_store_from_hdf_url (test helper) gained a drop_variables passthrough so the new tests follow the module's existing pattern.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.39%. Comparing base (117acc7) to head (e0e3fb5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1058   +/-   ##
=======================================
  Coverage   90.39%   90.39%           
=======================================
  Files          37       37           
  Lines        2248     2248           
=======================================
  Hits         2032     2032           
  Misses        216      216           

Files with missing lines	Coverage Δ
virtualizarr/parsers/hdf/hdf.py	97.05% <100.00%> (ø)

🚀 New features to boost your workflow:

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