ehrdata_blobs stores the time series in the default 3D .X

[sueoglu]

Closes #284

ehrdata_blobs was the last dataset loader still defaulting its tensor to .layers["tem_data"]. It now takes layer: str | None = None and puts the time series in the default 3D .X, matching physionet2012 / physionet2019.

What changed

The 2D static X is gone. The generator built the tensor first, then overwrote X with a per-observation mid-timepoint slice of it. Verified before removal, the old X was reproducible from the tensor to max abs diff: 0.0.

Docstring fixes. layer was documented twice in the Args block, and the first entry claimed "If not specified, uses X", the opposite of what the old default did. Collapsed to one accurate entry. The intermediate X variable is renamed base_values, since nothing named X is returned any more and the old name actively misleads about what .X holds.

Also corrected the to_pandas docstring, which advertised shape of .X: (2, 2) plus a .R field that no longer exists.

Requires anndata >=0.13 to use the default

ehrdata_blobs() with defaults now raises on anndata<0.13, which rejects a >2D X at construction:

ValueError: X needs to be 2-dimensional, not 3-dimensional.

On older anndata the time series has to go in a layer, so ehrdata_blobs(layer="tem_data", ...) remains the way to use it there.

Tests
tests/dt/test_dt.py updated: assertions read the tensor from .X, and a new test_ehrdata_blobs_layer covers the named-layer path, which had no coverage before, it asserts X is None and that both paths produce an identical tensor.

Two assertion groups were deleted rather than adapted, because they pinned the snapshot relationship this PR removes by design:

• the loop asserting edata.X[i, v] == edata.layers[...][i, v, mid_idx] for categorical variables
• the check that the tensor at t=0 correlates with X

[eroell]

This looks good!

If you merge main into this branch, the last CI check should also pass.

Please add an entry in the release notes (Claude sometimes creates a bit overly verbose release note comments, maybe shorten the first suggestion it makes), then this is ready to merge!