Add feature description for heterogeneous gpu inference feature (#707)

Signed-off-by: Varun Gupta <[email protected]>

README.md

@@ -12,6 +12,7 @@ The initial release includes the following key features:
 - **LLM App-Tailored Autoscaler**: Dynamically scale inference resources based on real-time demand.
 - **Unified AI Runtime**: A versatile sidecar enabling metric standardization, model downloading, and management.
 - **Distributed KV Cache**: Enables high-capacity, cross-engine KV reuse.
+- **Cost-efficient Heterogeneous Serving**: Enables mixed GPU inference to reduce costs with SLO guarantees.
 - **GPU Hardware Failure Detection (TBD)**: Proactive detection of GPU hardware issues.
 
 ## Architecture