refactor(io): serve MATLAB subscripts in the base lazy array

[ehennestad]

Each lazy array turns MATLAB subscripts (DataStub indexing, load_mat_style) into reads itself. The Zarr v3 lazy array in #853 re-implements the checks, expected-size rules and reordering of HDF5LazyArray, adds its own choice between a bounding-box read and element reads, and reverses dimensions in several places. This is the second point of the design note on #853.

io.backend.base.LazyArray now implements load_mat_style. It checks the subscripts, reads each dimension's indices once and sorted, and returns the values in the order and shape the subscripts ask for, as a table for a compound dataset. A backend implements two reads in its own storage order: readSubscripts reads every combination of per-dimension indices, and readPoints reads single elements. A linear selection is read through its bounding box or element by element, within limits the backend sets in getBoundingBoxLimits. A backend that stores arrays in row-major order returns true from isStorageOrderReversed, and the base class reverses subscripts and dimensions for it. readHyperslab serves start, count and stride through the same path.

HDF5LazyArray implements the two reads with its existing selections: one union of hyperslabs for subscripts and one element selection for points. The subscript rules and results are unchanged. A linear selection that fills its bounding box is now read as one hyperslab in any rank; before, only a contiguous run in a one-dimensional dataset was.

tests.unit.io.backend.base.LazyArraySelectionTest runs selections through an in-memory lazy array in both storage orders and compares them with MATLAB indexing of the same array.

This conflicts with #923, which restructures HDF5LazyArray/load_mat_style.m. That file's read code now lives in readSubscripts, which is where #923's grouped reads belong; the reshaping it moves into reshapeLoadedData is now in the base class.

🤖 Generated with Claude Code