Optimize rendering concurrency

[petrsloup]

Have more mbgl.Map instances ready to be able to process more requests simultaneously.

[petrsloup]

Dynamic pool of map instances (creating and releasing instances as needed) is implemented in mappool branch, but there are probably some issues in mbgl.Map itself, causing segfaults.

[petrsloup]

This is now in master.

There are some issues with the mesa3d graphics library and concurrent rendering.
The Dockerfile is now based on debian:stretch (testing) to get libgl1-mesa-dri 11.1.2 which does seem to be stable (as opposed to 10.3.2 in jessie).
Concurrent rendering (via multiple instances) is fully supported by mapbox-gl-native (see mapbox/mapbox-gl-native#4215).

The tileserver-gl server now has dynamic pool of instances with 4-16 instances for scale 1; 2-8 for @2x; 2-4 for @3x. (for each style)
It might be useful to make this somehow configurable in the config file (together with the available scales? -- we can easily go higher than @3x).

[klokan]

BTW our dedicated servers have hardware accelerated GPU with available X.org driver.
Before production deploy it would be worth to check performance if these are turned on vs. emulated on CPU.
Related ticket: https://github.com/klokan/internal/issues/39