[codex] Introduce LazyArray abstraction for HDF5-backed DataStub

[ehennestad]

Summary

This stacked PR implements the milestone 2 lazy-read seam on top of refactor-hdf5-backend-reader.

It introduces a backend-aware LazyArray abstraction for lazy dataset access, adds an HDF5 implementation, and refactors DataStub to delegate lazy reads through that abstraction while keeping scope HDF5-only.

What Changed

• add io.backend.base.LazyArray
• add io.backend.hdf5.HDF5LazyArray
• refactor types.untyped.DataStub to delegate metadata lookup and lazy loads through a backend lazy array
• update io.backend.hdf5.HDF5Reader to wrap non-scalar non-empty datasets in DataStub backed by HDF5LazyArray
• extend backend tests and add focused HDF5LazyArray coverage

Why

DataStub previously embedded HDF5-specific dataset metadata and read logic directly. That made lazy reads another HDF5-hardcoded seam in the Phase 1 architecture.

This change moves that responsibility behind a backend abstraction so the public lazy dataset proxy remains stable while storage-specific behavior lives in backend code.

Impact

• no intended user-facing behavior change for HDF5 reads
• DataStub remains the public lazy dataset object
• lazy dataset behavior is now backend-aware, which unblocks future backend expansion without another cross-cutting DataStub refactor

Validation

• tests.unit.io.backend.BackendFactoryTest
• tests.unit.io.backend.HDF5LazyArrayTest
• tests.unit.io.backend.HDF5ReaderTest
• tests.unit.dataStubTest
• tests.unit.nwbReadTest
• tests.unit.io.ParseDatasetTest
• tests.unit.dataPipeTest