paula: clock the POTxDAT counters at H-sync (once per scanline)

[LinuxJedi]

Problem

Denise clocks the POTxDAT counters once per horizontal line, at H-sync -- the
same per-line clock as CIA-B's TOD. Copperline advanced them on a free-running
512-colour-clock divider (~2.25 scanlines/step): the counting ran ~2.25x too
slow and was not phase-aligned to the beam, so the pot-counter register
visualisations came out with the wrong band spacing and phase.

Fix

• Drive the pot scan from the bus's per-new-line loop (next to the CIA-B TOD
  tick), tick_pot_hsync(), instead of a colour-clock accumulator.
• Wrap the 8-bit counter at 256 instead of saturating (a floating pin keeps
  cycling for as long as the scan runs).
• Keep the existing "a pin driven HIGH as an output freezes at 0" gate -- the
  Gods / Bitmap Brothers controls rely on it (pot_pin_driven_high).
• While a scan is running, cap the advance at the next line boundary (the same
  deadline CIA-B TOD uses) so a CPU read of POTxDAT lands on the current value;
  when no scan is running the pots impose no deadline (so idle software is
  unaffected).

Validation

• vAmigaTS Mainboard/Ports/POT: summed divergence 805.7% -> 790.8%.
  pot0dat3 becomes pixel-exact (0%); pot0dat4 30->20, pot0dat6 30->10,
  stop2 15->5, pot0dat5 10->5, cmp* slightly better.
• Known regressions (accepted): start1/2 15->25, potout1/2/3/4 by ~5% each.
  These are read-once tests gated on a vblank IRQ, so their value is the number
  of lines between the POTGO START and the IRQ read. The now-correct per-line
  counter is more sensitive to Copperline's interrupt/read beam-timing
  residual (the separate CPU/IRQ write-timing model) that the coarse 512-divider
  had masked -- they are "differently wrong", not "correct -> broken", and will
  resolve when that timing model lands. pot0dat1/cmp stay ~100% for the same
  reason (fully interrupt-timing-bound).
• Byte-identity: phooey / tomato / roots-ecs / roots-aga / zool screenshots
  stay byte-identical at 5s and 9s (the change only affects software that reads
  POTxDAT).
• 1303 unit tests green (POT tests rewritten for the per-line model); clippy and
  fmt clean.

Verified with the new tools/adf-compare.sh against a local vAmiga 4.4 headless
build.