test: fix test_dunedin_pace_normalization on pandas 3.0 (fixture orientation)

[IsraelAfangideh]

Summary

Fixes #209.

test_dunedin_pace_normalization fails on pandas 3.0 with
KeyError: 'cg00000029'. The DunedinPACE normalization output is correct —
the problem is that the test compares against a reference fixture in a different
axis orientation, and the comparison logic depended on stack()'s NaN handling.

Root cause

• actual (live loader output) is CpG-indexed (index = cg00000029, …;
  columns = sample GSMs).
• expected (testset/pace_normalized.pkl) stores probe IDs in an ID_REF
  column with a default integer RangeIndex.

The axes never aligned. On older pandas, DataFrame.stack() silently dropped
the misaligned NaNs, so actual[mask].stack() was empty and the test passed
without ever comparing values. On pandas 3.0 the NaNs are kept, so the
diagnostic loop runs and does a label-based .at[('cg00000029', …)] lookup
against the RangeIndex, raising KeyError. (This only surfaces once the
read-only-array fix from #206 lets normalization run to completion; before that,
it aborted earlier with a ValueError.)

Fix

• Align the fixture's orientation to actual before comparing (promote ID_REF
  to the index and reindex to actual's axes).
• Count mismatches from the boolean mask directly, so the result no longer
  depends on version-specific stack() NaN handling.

The normalization computation is untouched — after realignment the values match
the fixture to ~5e-16 (0 of 200000 cells exceed the 1e-6 tolerance).

Testing

• pytest biolearn/test/test_model.py::test_dunedin_pace_normalization — now
  passes on Python 3.11 / pandas 3.0.3 / numpy 2.4.6 (previously errored).
• Full test_model.py + test_dunedin_pace.py green on the same stack, apart
  from pre-existing, unrelated test_models[GPAge*] failures (missing optional
  GPy dependency in my local env).
• black --check clean.

Note

This is the test-side fix I offered in #209, chosen because it's fully reviewable
in plain code and keeps the existing fixture. If you'd rather regenerate
pace_normalized.pkl in the current CpG-indexed orientation instead, happy to
switch to that — just let me know.

[marcbal77]

Confirmed the fix locally on pandas 2.2.3 too. The realignment is genuine, 20000/20000 index overlap, 0 NaNs, max diff ~5.5e-16 across all 200k cells, so the values are actually being compared now instead of the old vacuous pass. Merging. This also covers #207. Picking up #208 next. Thanks for the contribution.