initial attempt at per package layering

[tofay]

As discussed in #153

implement per-package layering a la https://grahamc.com/blog/nix-and-layered-docker-images/, using nix's existing popularity algorithm.

The package graph is determine heuristically, by looking for the first installed package that satisfies each of a packages requires.
Cycles are then removed arbitrarily using Python's TopologicalSorter before the graph is passed to the nix popularity function.
This is likely to affect determinism of the image manifest:

• the graph we build is dependent on the results of dnf queries
• the cycle removal isn't guaranteed to be deterministic
• the nix popularity function results need sorting by package name

A layer is created for each of the most popular packages. Each layer has it's own "clamp_mtime", which is taken from the package build date, so package layer digests are consistent across builds.

TODO

• determine how configurable this should be (on/off, number of layers, batched layering??)
• investigate impact on per image reproducibility. It seems unlikely that
• sort the results from nix's popularity algorithm by popularity and name
• testing