Add Zarr archive and specification for ViT-MNIST inference traces (issue 39)

[Rosette2237]

Summary

This PR adds the Zarr archive design and implementation for storing ViT inference traces, including inputs, predictions, metadata, and per-layer attention outputs. The goal is to support efficient partial reads for downstream visualization and analysis, instead of relying on a simple .pt archive.

For the full archive specification, please refer to Zarr archive specification.

Included in this PR

• Zarr archive design for ViT attention data
• Archive structure for metadata, inputs, and per-layer attention outputs
• Storage choices for chunking and compression
• Supporting specification documentation in ARCHIVE_SPEC.md

Why

This design makes it easier to:

• read specific samples, layers, or heads without loading the full archive
• support attention visualization workflows
• extend the archive format for larger future models and datasets

Future Work

Design the similar Zarr archive for VizFold and fix issue #39.