brainglobe-cellfinder: expand input support to 2.5D and single-channel data (Soumya Snigdha Kundu)

Cellfinder currently expects full 3D whole-brain volumes with both signal and background channels, so brain slices (2.5D) and single-channel acquisitions are not first-class citizens. The codebase also mixes axis conventions, which makes dimensionality changes error-prone. A recent data-loading refactor introduces explicit axis tracking and a new class hierarchy for cube generation and augmentation; the 2D/2.5D and single-channel work proposed here will build on that foundation. The core deliverables are to standardise axis semantics as the foundation, add a 2D detection path for brain slices and refactor classification to support both 2D and 3D inputs, and make the background channel optional across the main API and classification pipeline.