fix: clamp HDF chunk dimensions to >= 1 for zero-length datasets (#977)

Zarr v3 allows `shape=(0,)` but forbids zero-length chunk dimensions,
now enforced by zarr-python 3.2.0+. When falling back to `dataset.shape`
for chunking, clamp each dim to a minimum of 1 so the HDFParser handles
empty HDF5 datasets across all supported zarr versions.

Author: TomNicholas

docs/releases.md

@@ -8,6 +8,9 @@
 
 ### Bug fixes
 
+- Fix `HDFParser` failing on HDF5 datasets with a zero-length dimension under `zarr-python >= 3.2.0`, which forbids zero-length chunk dimensions. Chunk dimensions are now clamped to a minimum of 1 when falling back from dataset shape.
+  By [Tom Nicholas](https://github.com/TomNicholas).
+
 ### Documentation
 
 ### Internal changes

virtualizarr/parsers/hdf/hdf.py

@@ -56,7 +56,10 @@ def _construct_manifest_array(
     -------
     ManifestArray
     """
-    chunks = dataset.chunks or dataset.shape
+    # Clamp each dim to >= 1: zarr v3 allows shape=(0,) but forbids zero-length
+    # chunk dimensions (enforced by zarr-python >= 3.2.0). See
+    # https://github.com/zarr-developers/zarr-python/issues/3711.
+    chunks = dataset.chunks or tuple(max(s, 1) for s in dataset.shape)
     codecs = codecs_from_dataset(dataset)
     attrs = _extract_attrs(dataset)
     dtype = dataset.dtype