The screen real-estate is already reserved, the values are dumped and
restored by the on-DSP code, why not make something out of these values ?
Allows following:
- where exactly send_back was called from ($st1)
- the boundaries and progress of the innermost BLOOP{,I} ($st0, 2 and 3)
  up to send_back's call

docs/DSP: Add sections on 16-bit and 40-bit modes and on main and extended opcode writing to the same register

The current code expects new mail almost immediately after the last map was sent for it to be saved properly. However, I have a test program that ends up looping for 32768 iterations before it sends more mail; this resulted in an incomplete result dump. I've changed it to wait a frame between checking for mail, which solves that issue. This does slow down dumping, but the end speed matches the speed at which the UI updates the registers so this isn't a big deal (the UI waits a frame between mail normally). (Theoretically, it could take even longer for dumping to finish, so this is not a perfect solution. However, for tests that take that long to run, it would be better to save the existing results instead of re-running the test and saving that; that'd be something to do with later improvements.)

DSPAssembler: Rework errors and warnings, and related cleanup

The assembler upper-cases the mnemonic internally, so it never would match the lower-case x.