feat: enable parallel e2e test execution for faster local dev feedback (#2785)

Run standard and experimental e2e tests simultaneously on separate KIND
clusters via `make -j2 test-e2e test-experimental-e2e`, cutting local
e2e wall-clock time by ~37% (25m → 16m). This is especially valuable
when working with AI coding agents that benefit from faster feedback
loops.

Key changes:
- Use Make pattern rules (kind-cluster-%, kind-load-%, etc.) with the
  stem identifying the variant. Cluster names, kubeconfig paths, and
  coverage names are derived by convention from the stem. Only
  variant-specific deviations need explicit overrides.
- Remove .NOTPARALLEL directive (no longer needed with unique targets)
- Isolate per-cluster kubeconfig files under .kubeconfig/
- Use separate KIND cluster names and prometheus host ports (30900 vs
  30901) to avoid resource conflicts
- Make PROMETHEUS_URL configurable via env var in summary reporter;
  skip summary generation when unset
- Conditionally serialize kind-deploy when both targets run together
  to avoid races on shared files

Linux prerequisite for parallel runs:
  sudo sysctl fs.inotify.max_user_instances=512

Co-authored-by: Claude <noreply@anthropic.com>