feat(vllm-performance): Added logic for K8s environments garbage collection

[christian-pinto]

Summary

Adds idle-environment garbage collection (GC) to the vllm_performance actuator. Previously, Kubernetes deployments released by an experiment were kept alive indefinitely until they were needed again or cleaned up manually. This PR introduces a configurable TTL-based free pool and a background GC loop that automatically deletes stale deployments, with an optional force-delete escalation path for stuck environments.

High-Level Changes

• Free environment pool with TTL — finished deployments are returned to a pool and reused for compatible experiments. A configurable free_environment_ttl (default 300 s) controls how long an idle deployment may live before GC deletes it. Setting TTL to 0 disables the pool and deletes immediately on release.
• Background GC loop — a new async loop runs periodically, deleting stale free environments and confirming that previously deleted deployments have actually disappeared from Kubernetes.
• Force-delete escalation — two new parameters (gc_force_delete, gc_force_delete_threshold) allow automatic force-deletion (clearing finalizers, grace_period=0) of deployments that remain stuck in Kubernetes after repeated GC checks.
• New K8s operation — ComponentsManager gains a force_delete_environment method that deletes the service, clears pod finalizers, and force-deletes pods and the deployment.
• New parameters — free_environment_ttl, gc_force_delete, and gc_force_delete_threshold added to VLLMPerformanceTestParameters with validation.
• Documentation — user guide updated with a new "Environment free pool and garbage collection" section, parameter descriptions, and configuration examples.
• Tests — new test file covers the GC loop, TTL behaviour, force-delete escalation, and K8s error handling.

Impact

• Cluster resource management — idle deployments are now automatically cleaned up, reducing unnecessary cluster resource consumption.
• Backwards compatible — existing configurations are unaffected; defaults preserve the previous behaviour of keeping deployments alive (TTL=300 s), though deployments will now eventually be garbage-collected rather than held indefinitely.
• Plugin version bump — 1.10.3 → 1.10.4 (patch release).

This PR fixes #273