feat(config): migration tool for legacy dataset identifiers → source: (M6c)

[wprazuch]

What

Adds nemo_gym.dataset_source_migration — a tool that rewrites the deprecated
gitlab_identifier: / huggingface_identifier: blocks to the unified source: block
introduced in #1637. Completes the migration-script deliverable of M6c.

# preview (no writes; exits non-zero if anything would change)
python -m nemo_gym.dataset_source_migration --dry-run resources_servers responses_api_agents

# rewrite in place
python -m nemo_gym.dataset_source_migration resources_servers responses_api_agents

How

A line-based transform, not a YAML round-trip: it matches a bare block-style key
line and rewrites only that line — gitlab_identifier: → source: + type: gitlab.
The already-nested fields (dataset_name/version/artifact_fpath, or repo_id) sit at
exactly the indent source:'s children need, so they're left untouched. Result:

• comments, key ordering, and formatting are preserved byte-for-byte;
• diffs are minimal (two changed lines per block);
• no new dependency (ruamel isn't available in the env).

Flow-style mappings and key lines with trailing comments are deliberately not matched
(rare here; rewriting them blindly risks corruption) — they're left as-is and reported as
0 migrations.

Why python -m, not an ng_* console script

The ng_* scripts run under the Hydra harness (and test_pyproject_scripts calls each
with +h=true expecting SystemExit), which a path-taking argparse migration tool doesn't
fit. A python -m entrypoint is the right shape and keeps this off the unified-CLI surface.

Validation

• Unit tests (100% coverage of the module): both backends, multi-block files, no-op files,
  flow-style/trailing-comment guards, dry-run vs in-place, and a round-trip test proving the
  migrated YAML validates as a DatasetConfig with no DeprecationWarning.
• Dry-run on the live tree reports 109 blocks across 59 configs — those can be migrated in
  a follow-up sweep PR (kept separate to avoid a 59-file diff colliding with in-flight branches).

Part of the configuration-friction epic (#1205), milestone M6c. Targets
wprazuch/dataset-source since it produces the source: schema from that PR.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.