Increase miner sampling for each step

[bkb2135]

Sampling 5% of the network every step introduces a lot of noise into the ema calculation, which leads to unstable incentives. This problem is mitigated by sampling all uids each step.

This PR has currently been run on testnet, but needs a mainnet run to examine stability, as this could potentially be a large change.

[Unkownman086]

Do you really think making 1024 connections is good idea? I believe default ubuntu limit is 1024 file descriptors if not modified.

I believe querying all endpoints in unnecessary

[bkb2135]

I believe default ubuntu limit is 1024 file descriptors if not modified.

In practice we should only query 1001 miners, because validators do not query each other. The rationale for querying all miners is that when we query only a small subset of miners, because of the different distributions of rewards on tasks, 2 models running the exact same code with the exact same intelligence can be on opposite ends of the incentive distribution based on who gets lucky with getting which queries, which does not incentivize intelligence.