External GPU support for big models

[adeebshihadeh]

USB GPU issues:

• compile is slow
  • time USBGPU=1 scons -u -j1 . takes 8.5m
• takes 2m to load the model on modeld init
• can't build all the models in parallel due to separate build processes?
• backend selection is determined at compile time and at runtime. put it in the pickle?
• getting random restarts on device during compile and modeld killing. no pstore was left behind though
• saturates the CPU core. busy waiting on something?

tinygrad regressions

• PC compile is broken?
• QCOM got a bit slower and now fails CI; can't bump until this is fixed