[kfox]

Adds tools/bench/regsweep.py and its C64 side, tools/bench/regread.a. Together they measure the NTSC/PAL read margin of whichever Special IO handler is chosen in Settings, as docs/Architecture/Constraints.md asks for a new handler. They are the tools behind the margin figures in #81.

How it works

• regread.a finds the cartridge's layout: plain REU, REU with the TR page at $DE00, or GeoRAM with the TR page at $DF00.
• It then loops, counting wrong answers at $C000:
  • the REU's address and length registers, read back as written;
  • $DF10-$DF1F reading $FF beside a TR page at $DE00;
  • the TR page's TRM/TRB IDs.
• regsweep.py DWELL VALUE... builds and launches the program, sets nS_PLAprop with tp### for each value, and prints the errors counted in DWELL seconds. It restores the defaults with td at the end.
• It needs acme on PATH and an image built with Dbg_SerTimChg, and refuses up front on one built without it.

Verified on a TR+ in an NTSC C128, using a Dbg_SerTimChg build of #81 under REU+USB Mouse/Disk:

• regsweep.py 4 150 240 250 260 reported the layout as "REU with the TR page at $DE00".
• There were no errors at 150 or 240. The first errors came at 250, which matches GEOS on the TR+'s REU: Special IO "REU+USB Mouse/Disk" #81's sweep.
• The bench README tests pass: 85 of 85.

Reviewed at medium, plus a low review of that review's fix commit.

• The medium review found that the error counts wrapped at 16 bits once most reads failed. On the board, $FF page printed +1248 at nS_PLAprop 400.
• The counters are now 24 bits, and the same run prints +66497.
• The low review reverted the third counter byte and saw the wrapped value come back, then restored it.