KEP-766: Add DisaggregatedSet proposal (#767)

* KEP-766: DisaggregatedSet proposal

DisaggregatedSet is a higher-level CRD that orchestrates multiple
LeaderWorkerSets for disaggregated inference workloads (e.g., prefill
and decode phases running on separate infrastructure).

Key features:
- Manages 2-10 phases as a single logical unit
- N-dimensional coordinated rolling updates using linear interpolation
- Auto-created headless Services per phase per revision for load
  balancer integration (e.g., llm-d can count pods per revision to
  route traffic proportionally)

* Update KEP-766: rename phase to role

Rename all API terminology from "phase" to "role" throughout the KEP:
- DisaggregatedPhaseSpec → DisaggregatedRoleSpec
- spec.phases → spec.roles
- disaggregatedset.x-k8s.io/phase → disaggregatedset.x-k8s.io/role
- Update all documentation references

The term "role" better describes the semantic meaning (prefill, decode, etc.)
and aligns with common usage in disaggregated inference literature.

Note: "phases of inference" in the motivation section is retained as it
refers to the actual LLM inference phases, not the API terminology.

* Add LeaderWorkerSetTemplateSpec type to KEP

Use LeaderWorkerSetTemplateSpec (similar to PodTemplateSpec) instead of
embedding LeaderWorkerSetSpec directly. This type needs to be added to
the LWS API and contains metadata + spec fields.