The speedup we might get is not too important, but the smaller binary
size is nice to have (currently, down from ~15 MiB to ~8 MiB).

We introduced the change to make MineMod compatible with mods that do
not have a mod.conf, but we forgot that read_conf()? will just make the
function return early in that case - and the fallback value is never
used.

This change makes it so that read_conf() is only called when a mod.conf
exists, which has the benefit that we get the legacy name resolution for
mods without mod.conf, but if there are I/O errors while reading the
config, we will still propagate them properly (same as dependencies()).

The user_choice() function was a bit weird because it needed a
"formatted" item for every item to select (not using the Display of the
item), but it also didn't support colored output. This change fixes that
by introducing a new trait that does colored output and by making
user_choice() use that trait to output the lines.

This allows mods to be installed via a git repository URL, such as
  git+https://gitlab.com/...

Key-Value-Stores are not always as easy as every line containing a
key-value pair, there is also the variant that a value spans multiple
lines and is enclosed in """ (e.g. the description of some mods).

This also changes the API slightly, as we want to separate the parsing
code from the code that reads the file. This makes it easier to do
testing or to re-use the code for data that is already in-memory.
Therefore, kvstore::write now takes a writer and kvstore::save provides
the functionality that the old kvstore::write had.

Those mods might not have a mod.conf and instead define their
dependencies in depends.txt. A lot of mods in mineclone2 for example
still do that.

Especially SubcommandRequired is helpful, as without one, we're not
doing all that much.