Distributed and federated modes should be just one networked mode

[Expro]

Is your feature request related to a problem? Please describe.

Currently, there are 2 separate networked modes - federated (load balanced) and distributed (llama-cpp distributed inference). This split is quite wasteful, as it forces user to setup separate load balanced cluster for non-llama-cpp workloads for HA, and then separate cluster for distributed inference for llama-cpp. Another issue is that switching between federated and distributed modes requires restart of entire cluster - it's startup defined.

Describe the solution you'd like

There should be just one networked mode that by default does load balancing for workloads, with toggle switching llama-cpp backend between load balancing and distributed inferencing.

This way:

• GPUs can be shared between backends even when llama-cpp is in distributed mode
• llama-cpp mode can be changed without restarting all nodes
• there is only one configuration path for networked mode - currently there are separate instructions for federated and distributed modes (and they are outdated)

I'm putting aside issue that currently distributed mode is broken and doesn't work at all.