Implement baseline ReflectionTable with eager HDF5 loading

[dimitrivlachos]

This PR introduces a canonical implementation of the ReflectionTable,
where all datasets are eagerly loaded into memory. While not optimized
for large-scale performance, this version provides a correct and
complete foundation for future enhancements such as lazy or chunked
loading. It simplifies the data model and establishes the foundation for
the public API and testing interface.

• Introduce a ReflectionTable class that eagerly loads HDF5 datasets
  into memory.
• Add support for extracting dataset shape via
  read_array_with_shape_from_h5_file().
• Refactor h5read_processed.hpp to support recursive group traversal,
  metadata extraction, and shape-aware reading.
• Add write_raw_data_to_h5_file() and related helpers in h5write.hpp
  for raw and high-level writing with shape/attribute support.
• Add h5dispatch.hpp to centralize C++/HDF5 type dispatch and
  registration.
• Integrate the mdspan backport for flexible memory views.
• Add DX2_DEBUG-controlled logging via utils.h.
• Add comprehensive unit tests covering filtering, writing, round-trip
  loading, and column type access.

[dimitrivlachos]

Yes, you are correct @jbeilstenedmands. I am currently working on this while integrating this PR into the Fast Feedback Service and will push those changes once they are working.

[jbeilstenedmands]

Thanks for undertaking this extensive piece of work, this looks really nice and functions well. I have verified that I can create and save h5 files that can be read by dials, so I think this can now be merged in.