Make the benchmark results more precise.

I have run a number of benchmarks by now. Results depend on a lot of
boundary conditions, sometimes in surprising ways. But I guess that's
true for any benchmark... Anyway, the text now gives at least one data
point to get an impression of what is doable.

Value is in the sample, not the labels.

Add note about memory usage during build.

Add Python client

add annotations documentation.

Fix regex, mention reserved labels.

Point out README.md build section from tutorial.

Add some FAQs.

- FAQ about timezone.
- FAQs about implementation.
- Fix swapped issue URLs.
- Add section about possible hash collisions.
- Mention `GOMAXPROCS`, especially important for Go beginners.
- Mention possible storage optimization for histograms.

Clarify that to_nearest may be a fraction in round().

Document rounding functions.

Add collectd. Reorder based on likely popularity.

Add staleness and clients/exporters

Add FAQ section on stability guarantees.

Improve documentation about storage.

Also add related FAQs (including the ZFS bug).

Add roadmap documentation.