This has the benefit that it removes the quotes, and it works better on
Windows, where double slashes were used.

As it turns out, the "local timestamp" as advertised by arcdps is a bit
misleading, because the timestamp is still in UTC. The "local" refers to
the fact that it can lag behind the server timestamp a bit (but usually
they seem to be within +-1 of each other), not that the timestamp is in
the local timezone.

This makes date handling a bit harder for raidgrep, but thanks to
chrono, not by much. The idea is that we simply deal with Utc pretty
much everywhere, except at the user boundary. This means that

1. Input timestamps for -before and -after are converted to Utc right
   after input
2. When outputting, we convert to a local timestamp first

This makes the output consistent with the filenames now (and the "wall
time" that the player saw).

It's kinda silly to have new() be generic when all it does is box the
objects anyway. It only makes code harder to write, as we cannot unify
the types to call Pipeline::new(), and instead we have to rely on
calling Pipeline::new() in the branches themselves.

It's not the biggest issue when we only have different formats, but at
some point we might want to add different aggregators as well (like a
sorting one or a counting one), so it would be bad if we suddenly had to
add all those branches.

This fix changes that, and we can build the pipeline piecewise by
having a Box<dyn Format> around, allowing us to combine it freely with
any Box<dyn Aggregator>.

Filtering based on guilds is slow, as it will have to retrieve every
guild name from the GW2 API, and it has to parse every log file instead
of bailing early.

Therefore, guilds are not searched by default, and have to be
explicitely turned on with --guilds.

In addition, this means that raidgrep will now need network access when
--guilds is passed, which was not the case before.