Why are these changes needed?

KubeRay's health checks and the RayJob submitter's GCS wait loop used to rely on wget to perform HTTP requests. Slim Ray images (Ray 2.53+) do not include wget, which led to probe or job submission failures. These changes remove the wget dependency for Ray ≥ 2.53 so KubeRay works with slim images that only provide Ray (and Python). For Ray < 2.53, wget is still used where exec probes or the GCS wait are needed, since slim images are only released for 2.53+.

Change summary

• RayCluster (head/worker): For Ray ≥ 2.53, use HTTP probes to /api/healthz (no wget). For Ray < 2.53, keep wget for exec-based health checks to the same endpoints.
• RayJob submitter (K8s Job / Sidecar mode): For Ray ≥ 2.53, the "wait for GCS" loop uses a Python-based health check instead of wget. For Ray < 2.53, the existing wget-based wait is kept.
• RayService worker: Readiness uses a single HTTP probe to /-/healthz on the serving port (Serve proxy health). Liveness uses node health (/api/healthz) to check.

Related issue number

Closes #3837

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(