Clearing the current active OpenGL context is apparently very expensive. Only do it if we're destroying an OpenGL context.

Fix the error handling: only create one call unless there's an error and only check for errors in functions that can fail

Avoid copying by passing matrices by reference to RenderDevice
Use fasttrig from GZDoom for faster cos/sin
Don't set matrices unless they changed. Even though the memcmp prevents it from being pushed to OpenGL it is still a waste.

Place all vertex buffers in the same GPU vertex buffer so that the 3D scene only contains 2 vertex buffer binds per frame