[Umbrella] advanced traffic load balancing algorithms

[kerthcet]

What would you like to be added:

We will focus on LLM-specific characteristics to load-balance traffic, like prefix-cache aware, kv-cache aware, lora-aware, load-aware, request-profile aware(summary or chat) and so on.

They're plugins baked into the envoy gateway.

• random selection as template and baseline, Envoy gateway plugin support with random selection #371
• LoRA aware plugin
• Fairness sharing
  • https://arxiv.org/abs/2401.00588
  • https://arxiv.org/abs/2501.14312
  • https://arxiv.org/abs/2407.00023
  • https://arxiv.org/abs/2312.07104
• prefix cache aware plugin
  • https://arxiv.org/pdf/2503.16525
  • https://arxiv.org/pdf/2405.16444

Why is this needed:

Better performance.

Completion requirements:

This enhancement requires the following artifacts:

• Design doc
• API change
• Docs update

The artifacts should be linked in subsequent comments.