Whereas run.py runs piglit's shader_runner binary to compile each shader
individually and parses the output of INTEL_DEBUG=fs,vs,gs to find the
number of instructions and loops, this runner compiles all of the
shaders from a single process and uses output from GL_KHR_debug to get
the information we want.

It uses EGL and GBM (and render nodes) to create a GL display and uses
libepoxy for GL function pointer management.

It creates one thread per-CPU using OpenMP, each of which compiles
shaders in parallel. It creates two OpenGL contexts, one core context
and one compatibility context and switches between them as needed.

run.py is able to compile all of the GLSL shaders in shader-db
(including the closed portion) in about 300 seconds on my quad-core
Haswell. This program can do the same in 90 seconds.

Profiling shows that it's largely limited by malloc performance, and
preloading jemalloc (LD_PRELOAD=/usr/lib64/libjemalloc.so.1) reduces the
execution time to about 80 seconds.