[FEATURES] RBG YAML Deduplication & Simplification for RBG

[cheyang]

Problem Statement
When deploying large-scale inference services, we repeatedly hit the same pain points:

1. Config explosion – every role (prefill / decode) needs its own 200-line YAML.
2. Copy-paste errors – changing an image tag or env value across dev / staging / prod is manual and error-prone.
3. Cross-team friction – platform engineers want to roll out new base images, probes, or affinity rules without forcing every business team to re-apply YAML.
4. Leader vs Worker drift – the same template must express different startup flags, ports, or resource requests for leader and worker pods, which today requires two nearly-identical manifests.