Implement a simple example Load Balancer

[vidas]

Implemented example load balancer for NekkoAPI

This commit implements a simple load balancer
capable of (somewhat) efficiently routing
request to the best worker nodes in the k8s
cluster context.

Load balancer features:

• discovers worker nodes in a cluster based on app label
• performs reverse proxying of requests to the worker nodes
  (inclusing SSE responses)
• tracks "busy" status of all workers
• tracks simulated kv-cache state of all workers
• routes requests according to:
  1. model availability
  2. busy status
  3. (simulated) kv-cache match
• exposes /v1/models endpoint and lists all models

Missing/broken parts:

• model names are hardcoded until we have proper model
  registry and alias->disgest mapping (this
  will be implemented in the control plane)
• proper logging
• auth
• documentation
• tests
• good error handling
• everything else (a lot of it)

Add k8s manifests to run load balancer in nekko
example k8s cluster.

Cody is an ugly hack mostly intended
as a proof-of-concept. Good parts a to be reused in the
control plane.

Also, removed llama model from examples,
during development the smoller the better.

Additionaly this PR fixes threading/async bug:

Fix async/threading issues and refactor server endpoints.
(This part was mostly cherrypicked (with small adaptations)
from llama-cpp-python upstream. Thanks!)

Remove unused endpoints.

Remove request interruption feature completely.