Make a background persistence thread (write back)

[petethepig]

Currently the only time pyroscope saves data on disk is when it stops or when the cache is overflowing. On large data sets this takes a long time, sometimes over 10 seconds.

Most service managers, e.g systemd or supervisord have a stop timeout of about 10 seconds. This means that if the process takes more than 10 seconds to stop they SIGKILL the process, leading to data loss.

This problem will become even bigger once we implement #167 because on larger systems it will take even more time to finish writes.

We need to change our cache system in such a way that it would write to disk more often, before we stop the process.

I draw a little illustration for this:

---

It's also important to not write too much, because in that case we're risking overloading disk, which would not be great.

[petethepig]

We could also improve write performance with:

• concurrent writes (multiple goroutines)
• batched writes