Update algorithm for issuing image generation tasks

[gpavelgr]

Implement proper logic for issuing tasks for different nodes within the network relying on "our" network of gpu's, Vast.ai network and other providers. The algorithm should prioritize our network first and attempting to access other resources in case of too big load (can be measured in expected delay for receiving an image for submitted request, f.e. longer than 1-2minutes).

After sync on Friday:

• It looks more like the task of finding cab drivers in Uber
• Server must orchestrate models

Problem:
different models are optimized for different video cards of miners, if we do batch queries, we need to take this into account

[gpavelgr]

Question/decision that affects the task:
How important is the speed of response to the prompt to the user? In general and for the first time (mvp 1-2)

[Savokr]

Best solution for now: create a mocked node server that will emulate different pools of nodes (paid Vast.ai, Azure cloud, custom network with different GPUs).