fix(checkpoint): allow failed checkpoint cleanup

[spichen]

A checkpoint whose snapshot creation fails cannot be deleted: the store only admits READY and DELETING records, so the owner receives NotFound and the failed checkpoint keeps its runtime revision pinned.

Allow FAILED checkpoints to enter the existing deletion workflow. When Substrate confirms the tag is absent, remove the checkpoint record without another delete-by-name request. If a tag exists, require a retained UID and matching identity before deleting it. Backend failures preserve the record for retry, and ownership and fork-reference checks remain in place.

Refs #2746. This addresses failed-checkpoint deletion only; automatic recovery of abandoned CREATING/DELETING operations remains follow-up work.

Validation

• New database regression fails on unchanged main with begin delete Session checkpoint: record not found and passes with the fix.
• New service regressions fail on unchanged main and pass with the fix.
• go test -race ./core/internal/service/checkpoint -count=1 passes.
• go test -race ./core/internal/database -skip '^TestSessionExpiration' -count=1 passes, including inline SQL preparation and runtime-revision retention coverage.
• go test -race ./core/pkg/migrations -count=1 passes.
• Focused golangci-lint v2.13.0 reports zero issues for the database and checkpoint service packages; formatting and git diff --check pass.

The unfiltered database suite has failures in five session-expiration tests. The same failures were reproduced with the production code restored to the base main revision. Full repository lint was blocked during tool setup by local Go checksum-cache permissions; live Substrate E2E validation has not been run.

Tests were run against base revision f2d7cc4. The change also documents failed-checkpoint cleanup semantics.

AI assistance: Codex helped implement and validate this contribution.