This batches up writes to the filesystem, which is more efficient for
disk I/O. This means that it can take some time for logs to get written
to disk. Note that ERROR logs (and above) immediately flush the buffer.

This only effects new installs, as we only write the log config if
started with `--generate-config` (in the same way we do for generating
signing keys).

This hopefully reduces the amount of logs kept for new servers. Keeping
the last 1GB of logs is likely overkill for new servers, but equally may
not be enough for busy ones.

Instead, we keep the last four days worth of logs, enough so that admins
can investigate any problems that happened over e.g. a long weekend.

We do this to prevent foot guns. The default config uses a MemoryFilter,
but users are free to change to logging to files directly. If they do
then they have to ensure to set the `filters: [context]` on the right
handler, otherwise records get written with the wrong context.

Instead we move the logic to happen when we generate a record, which is
when we *log* rather than *handle*.

(It's possible to add filters to loggers in the config, however they
don't apply to descendant loggers and so they have to be manually set on
*every* logger used in the code base)

We don't WARN too frequently, so this is fine and ensures that we log
things like "STARTING" immediately.

This reverts commit c2fc81e.