Split context to improve compile times

[solidpixel]

The ParallelManager utility is only needed in the entry layer, but is included everywhere in the internal header. This can be slow to compile due to the amount of C++ stdlib it pulls in, and the use of templates prevents the use of pre-compiled headers.

The PR splits the codec context into two - an inner context which excludes the thread control, and an outer context which uses it. This improves compile times by ~2 seconds (~20%) on Linux. Compiling Debug builds all three x86-64 binaries with -j16 drops from ~15s to ~12s.