feat(conformance): capture observed state in evidence artifacts

[dims]

Summary

• Replace hardcoded congratulatory strings in conformance evidence artifacts with actual observed cluster state
• Each behavioral check now returns a typed report struct capturing real values (HPA desired/current replicas, node counts, scheduling timestamps, webhook rejection codes, etc.)
• Fixes TestSecureAcceleratorAccess flaky failure on GPU Inference CI (was timing out at ~309s, now completes in ~12s)
• Deduplicates HPA scaling intent poll into shared helper

Changes

• cluster_autoscaling_check.go — capture HPA desired/current, baseline/observed node counts, scheduled/total pods
• pod_autoscaling_check.go — capture HPA desired/current, scale-up/down deployment replicas
• robust_controller_check.go — capture webhook rejection HTTP code, reason, and message
• secure_access_check.go — capture isolation pod name, phase, exit code, resource claims count; pin no-claim pod to GPU node via NodeName; add stuck-pod detection and rate limiter handling
• gang_scheduling_check.go — capture scheduling timestamps, co-schedule span
• inference_gateway_check.go — capture listener routes, HTTPRoute count, endpoint counts, condition details
• ai_service_metrics_check.go — capture HTTP status, group version, API resource count
• helpers.go — add podStuckReason(), podWaitingStatus(), and shared waitForHPAScaleUp() helpers
• Unit tests updated to verify report struct values

Test plan

• go test -race ./pkg/validator/... passes locally
• CI: Unit, Integration, Conformance, E2E tests pass
• CI: GPU Training Test (nvkind + H100 x2) passes
• CI: GPU Inference Test (nvkind + H100) passes — TestSecureAcceleratorAccess completes in 12s

[dims]

Fixing comments from @mchmarny in #201 (comment)