feat(DIARCHERS-1498): add MCP restart/abort/rerun endpoints

[sap-yuan]

Summary

Implements DIARCHERS-1498 — exposes job restart, rerun, and abort capabilities via the MCP API namespace.

Depends on #620 (MCP traffic isolation — must merge first).

5 new endpoints

All require allow_trigger = TRUE on the MCP token, project scope check, rate limiting (trigger_build bucket, 5 RPM), and audit logging.

Endpoint	Description
POST /api/v1/mcp/projects/<id>/builds/<bid>/restart	Restart a build — re-queues Create Jobs and all subsequent jobs
DELETE /api/v1/mcp/projects/<id>/builds/<bid>/abort	Abort all running jobs in a build
POST /api/v1/mcp/projects/<id>/jobs/<jid>/restart	Restart a job and all its downstream dependents
POST /api/v1/mcp/projects/<id>/jobs/<jid>/rerun	Rerun a single job without touching downstream (most common AI use case)
DELETE /api/v1/mcp/projects/<id>/jobs/<jid>/abort	Abort a single running job

Key design notes

• restart_build and restart_job delegate to the same DB logic as the existing /api/v1/* handlers via a shared _clone_jobs() helper — no business logic duplication
• abort endpoints use get_mcp_user_id() instead of g.token['user']['id'] since MCP auth sets a different context
• rerun_job (single-job retry without downstream restart) is new — not currently in the regular /api/v1/* API as a separate verb, maps to the existing /rerun endpoint logic

Test plan

• restart_build creates a new build with incremented restart_counter
• abort_build inserts abort records for all jobs in the build
• restart_job clones job + all downstream dependents with .1 suffix
• rerun_job clones only the specified job
• abort_job inserts a single abort record
• All 5 return 403 when allow_trigger = FALSE on the token
• All 5 return 403 when project not in enabled_projects
• All 5 produce mcp_access_log audit entries