Bug: HDF5 variable attributes display as byte reprs (`b'NASA'`) in the comparison report

[danielfromearth]

Summary

Variable-level attributes read from HDF5 files are rendered as Python byte reprs
rather than text. An ATL06 comparison prints

long_name:   b'Signal Selection Source Fraction 3'
source:      b'ATL06 ATBD - Q/A Statistics'
units:       b'1'

where the equivalent netCDF attribute would show Signal Selection Source Fraction 3. On the two committed ATL06 structural fixtures there are 4232
such occurrences.

Beyond being ugly, it can manufacture a false difference: if two files store the
same attribute under different HDF5 string types (File A fixed-length, File B
variable-length), the values render as b'NASA' and NASA and are counted as a
difference. Confirmed by building exactly that pair — before decoding they are
reported as differing; with decoding the comparison returns a total of 0.

#77 (PR #344, merged) fixed this for global (root-level) attributes only.

Why the existing decode doesn't cover this

_value_to_comparable_str in getters.py decodes bytes correctly, and root
attributes reach it as raw h5py values. Variable attributes do not: the HDF5
branch of Comparison._create_var_properties stringifies eagerly as it reads,

else:
    retrieved_value = str(attribute_value)   # Comparison.py:563

so by the time get_attribute_value_as_str → _value_to_comparable_str sees the
value it is already the string "b'NASA'". isinstance(value, bytes) is False
and the decode is a no-op. The fix therefore belongs in
_create_var_properties, not in _value_to_comparable_str.

Proposed fix

Stop stringifying at read time, or decode before doing so, in the hdf5 branch of
_create_var_properties. Note there are three str() calls in that branch, not
one — the scalar case (:563), the list-of-references case (:556), and the
IndexError fallback (:560).

This is more delicate than the root-attribute change: the same branch resolves h5py
object references via __name_from_h5_ref, and h5py.ref_dtype values must keep
going down that path. The str() cannot simply be dropped. The cleanest shape is
probably to hand the raw value to _value_to_comparable_str for the non-reference
cases and leave reference handling untouched.

Verification

• A cross-string-type regression test: two HDF5 files storing the same variable
  attribute as fixed-length and variable-length strings should compare equal (this
  is the false-difference case above; the root-attribute equivalent is
  test_get_root_attributes_hdf5_fixed_length_string_matches_netcdf).
• EXPECTED_ATL06_DIFFERENCES will move again. It is 4978 as of Add global attributes to comparison display #344,
  merged in c2a18b4. Report the number the suite actually produces and state the arithmetic,
  as with Add global attributes to comparison display #344.
• The netCDF golden files should be unaffected, since netCDF4 already returns str
  for text attributes — worth confirming rather than assuming.

Context

Split out of #77 / PR #344 deliberately, to keep that PR's scope to root-level
attributes and avoid moving the pinned ATL06 count twice in one change.