Eliminate file I/O for draco encode step

[stuarteberg]

Currently, mesh_from_array() relies on two external command-line tools (fq-mesh-simplify and draco_encoder), both of which use filepaths for input and output (not streams). A naïve implementation would incur the cost of disk I/O several times:

1. mesh_from_array() writes the raw .obj to disk
2. fq-mesh-simplify reads the raw .obj from disk
3. fq-mesh-simplify writes the simplified .objto disk
4. draco_encode reads the simplified .obj from disk
5. draco_encode writes the compressed .drc to disk
6. mesh_from_array() reads the compressed .drc from disk

Fortunately, most of those I/O costs are avoided in mesh_from_array() by using unix "named pipes" to write and read the data. Unfortunately, steps 3. and 4. could not be avoided: draco_encode requires a seekable stream (for no good reason, IMHO), which means we're forced to use the hard disk.

The only way to eliminate that last I/O penalty is to implement our own version of draco_encode, or to implement python bindings for the draco::Encoder C++ class.