[amh1k]

Summary

Closes #56
Introduce a modular benchmark driver architecture and structured runner metrics, following the design discussion in #49.

Changes

• Add a driver registry and pgbench implementation that owns option defaults, strict validation, connection handling, and workload preparation.
• Accept explicit driver and nested options in run requests, retaining the legacy pgbench request format for compatibility.
• Keep Kubernetes Job/Secret creation, scheduling, log collection, and cleanup in the shared coordinator.
• Record the driver, runner image, and effective options in run snapshots.
• Emit versioned pgbench metrics through /dev/termination-log and return them alongside logs in run status responses.
• Preserve execution status when metrics are missing or invalid, filter sensitive metadata, and protect stored results with defensive copies.

Validation

• Backend race tests and go vet pass.
• Tests cover driver preparation, shared execution, metrics decoding, authorization, redaction, storage, and cleanup.
• Metrics remain available when logs are truncated.
• Live Kubernetes verification remains pending.

Follow-up

Helm driver configuration and frontend integration will be handled separately.