feat(ray-on-eks): add single-request latency benchmark for Gemma 4 12B

[vara-bonthu]

What does this PR do?

🛑 Please open an issue first to discuss any significant work and flesh out details/direction. When we triage the issues, we will add labels to the issue like "Enhancement", "Bug" which should indicate to you that this issue can be worked on and we are looking forward to your PR. We would hate for your time to be wasted.
Consult the CONTRIBUTING guide for submitting pull-requests.

Adds a benchmark harness that measures TTFT, end-to-end latency and decode throughput for the Gemma 4 12B endpoint. Runs inside the cluster so client network RTT doesn't skew TTFT, and makes prefix-cache state explicit (--cache-mode cold|warm) so repeats don't silently report cache hits.

Also fixes issues found while measuring:

• max_model_len 32K → 64K — 31K-token prompts left too little room for output
• Pin the GPU worker to on-demand; a spot reclaim forces a full cold start
• Disable Ray autoscaling — a lost worker was causing two GPU nodes to launch
• Keep the Ray head in the same AZ as the GPU worker
• measure-deployment.sh was misreporting warm starts as cold

Motivation

More

• Yes, I have tested the PR using my local account setup (Provide any test evidence report under Additional Notes)
• Mandatory for new blueprints. Yes, I have added a example to support my blueprint PR
• Mandatory for new blueprints. Yes, I have updated the website/docs or website/blog section for this feature
• Yes, I ran pre-commit run -a with this PR. Link for installing pre-commit locally

For Moderators

• E2E Test successfully complete before merge?

Additional Notes