[codex] Add retry-aware image Lance workflow tutorial

[VibhuJawa]

Summary

Adds a focused NeMo Curator image-to-Lance tutorial workflow for retry-aware WebDataset-style tar inputs.

• builds one immutable source manifest and deterministic snapshot ID
• uses stable (shard, tar_id) file partitions across attempts
• selects one URL winner by resolution, then newest attempt, then stable source order
• streams tar ranges through fsspec without staging or re-uploading source bytes
• writes image payloads as Arrow large_binary with MD5/SHA-256 and winner provenance
• packs work near 1 GiB and preserves logical Slurm-array shard identity for sparse retries
• uses matched one-CPU materialization/Lance writer stages so Ray can fuse them
• documents the retry locality requirement and Slurm-array submission flow

Why

The image corpus contains repeated download attempts and needs deterministic, resumable distributed ingestion. A stable manifest and partition contract let Slurm arrays and Curator checkpoints retry exactly the failed logical work while keeping URL winner selection local to each stable tar partition.

Scope

This PR intentionally contains only the reusable core workflow and tutorial. Production-specific validation, cleanup, benchmarks, compaction, indexing, credentials, paths, and dataset state are excluded.

Validation

• Ruff format check
• Ruff lint
• 4 tutorial workflow tests
• Python compileall
• shell syntax check
• CLI help smoke test
• git diff --check