Compute better scoring for node performance

[jorgeantonio21]

Following issue atoma-network/atoma-node#410, we need to improve the node performance scoring computation, through by incorporating these more fine grained collected metrics. Ideally:

• Node specs are used to compute scores. These include KV cache allocations, thermal power values, cpu/gpu usage, bandwidth capacity.

These scores are used by the proxy to select the best nodes available to newly purchased Stacks (ahead of time). Then the proxy can use metrics around usage and performance, such as:

• Number of allocated KV blocks, node processed requests, throughput and latency (ITL and TTFT) across the board, to pick the most available node to process a new request (real time).