The previous implementation of `snrt_dma_wait` was incorrect as it had an infinite loop in the case that `tid` was the last completed transaction. Looking at the documentation, the snippet in custom_instructions.md shows the correct way of implementing the transaction check (get last completed tid and loop if it's less than the one we are waiting for) except it had its register operands swapped.
This PR uses that implementation instead, fixes the small bug in the documentation and changes one of the waits in `dma_simple.c` to make sure we don't regress either.