Add server-side Slurm-job correlation endpoint

[daniel-thom]

Summary

Tier 2 follow-up to the torc status server-side migration (#363). torc slurm diagnose-logs correlates Slurm job IDs to the Torc jobs they ran. It did this in build_slurm_to_jobs_map by fetching four full lists — scheduled compute nodes, compute nodes, results, and jobs — and joining them through three HashMaps in memory. This was the heaviest client-side join in the codebase.

This PR adds GET /workflows/{id}/slurm_job_correlations, which performs the entire join in a single SQL query:

scheduled_compute_node (scheduler_id = Slurm job ID)
  → compute_node      (linked via the scheduler JSON's scheduler_id)
  → result            (compute_node_id)
  → job

grouped/ordered by (slurm_job_id, job_id) and covering all runs (matching the prior all_runs=true behavior). Each table is narrowed by its workflow_id index before joining — EXPLAIN QUERY PLAN shows it starting from result via idx_result_workflow_id then primary-key lookups, with no table scans.

build_slurm_to_jobs_map now makes a single call and rebuilds the same HashMap<String, Vec<AffectedJob>>, so all consumers (diagnose-logs output) are unchanged. The now-unused pagination imports are removed.

Response shape

// GET /workflows/{id}/slurm_job_correlations
{ "items": [ { "slurm_job_id": "987654", "job_id": 42, "job_name": "train" } ] }

The client groups items by slurm_job_id; the server already deduplicates (GROUP BY) and orders the rows.

Testing

• New integration tests: test_get_slurm_job_correlations (builds the full SCN→compute_node→result→job chain and asserts the correlation) and test_get_slurm_job_correlations_not_found
• cargo fmt --check, cargo clippy --all --all-targets --all-features -- -D warnings, dprint check — clean (pre-commit hook)
• OpenAPI codegen parity check — ok; Rust/Python/Julia clients regenerated

🤖 Generated with Claude Code

[daniel-thom]

Superseded by #366, which consolidates all the remaining server-side aggregation work (these two commits are the first two there) plus the running-jobs command and the results/compute-node listing improvements into a single PR.