Idea: `deployment.evalOnTarget`

[mrshmllow]

If --path is given a pure flake ref, it would be possible for the node to evaluate itself.

In the case of most remote deployments, the link speed between your machine and the node can be quite slow. Pushing an evaluation can upload a tonn of data. This would solve that problem.