This allows the compile stage to optimize most of the translate()
function away and saves a ton of space (~40k on ESP). *However*, it
requires us to wait for the qstr output before we compile the rest
of our .o files. (Only qstr.o used to wait.)
This isn't as good as the current setup with LTO though. Trinket M0
loses <1k with this setup.
So, we should probably conditionalize this along with LTO.
