Initial E/PD extension for kind deployment

[revit13]

This PR is part of the implementation for #608. It extends the kind deployment to support Encode disaggregation.

This update configures the deployment to spin up a disaggregated architecture consisting of:

• 1 Encoder pod

• 1 Prefill/Decode (PD) pod

Notes:

• vLLM Instance: This setup currently provisions a real vLLM instance instead of a simulated one. Simulator support for this architecture will be added in a future update.
• Testing: This deployment configuration was actively used to validate the changes introduced in initial E/PD extension of the sidecar #643 and Initial E/PD extension for plugins #647.

Usage:

export EPD_ENABLED=true
export VLLM_SIMULATOR_IMAGE=ghcr.io/revit13/vllm-cpu-env:latest
export MODEL_NAME=Qwen/Qwen3-VL-2B-Instruct
make env-dev-kind