TVM and custom accelerators

@areusch for step 2, if you also wanted to use TVM for the accelerator codegen as well, would VTA provide a guide there? And would you still use BYOC there (just calling out to TVM in BYOC for compilation as well)?