dataset rm: remove the central backend catalog entry (server-side delete path)

[aptracebloc]

Follow-up from #30 (dataset rm). Spun out so the CLI-direct teardown can ship without blocking on cross-repo work.

Problem

A dataset push creates three artifacts:

1. the MySQL table (training_test_datasets.<table>),
2. the dataset's files on the shared PVC, and
3. a central tracebloc backend catalog entry (the "edge label metadata" the in-cluster ingestor syncs to the backend after a successful ingest).

dataset rm (#30) removes (1) and (2) in-cluster via exec — but the CLI has no direct line to the central backend, so (3) is left orphaned: a catalog entry pointing at a table/files that no longer exist.

This is observable today: the two datasets that fully ingested during v0.1 testing (clitest_reg_train, clitest_tc_train) reported records "sent to API", i.e. they created catalog entries; the manual cleanup dropped the in-cluster artifacts but not those entries.

Options

• jobs-manager delete-ingestion endpoint (mirrors submit-ingestion-run) that removes all three server-side, including the backend call. Cleanest; needs work in tracebloc/client + data-ingestors.
• A backend delete API the CLI could call directly (if one exists / is acceptable to expose).

Acceptance

• dataset rm (or its server-side counterpart) removes the central catalog entry for a successfully-ingested dataset, leaving no orphan.
• Decide the architecture with the team lead (the open question noted on feat(dataset): delete pushed data (dataset rm) [needs design] #30).

Refs #30, #31.

[aptracebloc]

Resolved — the tracebloc backend now removes a deleted dataset's catalog metadata automatically after deletion, so there's no CLI-side catalog teardown to build (and the CLI has no direct backend line anyway, which is why this was split out).

The CLI communicates the new behavior on branch feat/feedback-messages:

• dataset rm success output: "The dataset's catalog metadata will be removed automatically — no further action needed."
• dataset rm --help NOTE updated to match (replacing the old "NOT removed" caveat that referenced this issue).

Closing as the catalog cleanup is handled server-side.

— drafted with Claude (Opus 4.8), sent by @aptracebloc