This avoids multiplying the RTM compilation time by the number of AMC packages.
This also makes them a single artiq_flash target, and a single conda package.