denise: move the DIW comparator anchor to the hardware 2H-196 mapping

[LinuxJedi]

Summary

Move the Denise display-window comparator anchor one lores pixel left:
DIWSTRT/DIWSTOP hstart H now maps to framebuffer x = 2H - 196 (was 2H - 194).
Sprites follow (they share Denise's counter); bitplane fetch references move
+1 in lockstep so bitmap pixels keep their hardware-calibrated beam-anchored
positions. Net effect: every display window edge and sprite moves 2
framebuffer px left; bitmaps and colour writes stay.

This is the second half of the 1-lores-px divergence behind issues #68/#69
(the first half, early-DDF picture phase, landed as #10x). A standard lores
screen now shows its first bitmap sample flush at the window edge, as real
hardware does.

Hardware arbitration

vAmigaTS Agnus/Blitter/sblit/sblit0 (lores, DDF $38, standard DIW, image
flush to both window edges; real photo sblit0_A500_ECS.jpeg). In the
emulators, interior content positions agree exactly; the only difference is
the left window clip: vAmiga shows the leftmost partial swatch column 24 fb
px wide (= the right one), Copperline clipped it to 22. The photo (subpixel
edge detection, ~100 rows averaged per band, widths normalized by the
16-fb-px inter-swatch gap) measures the left partial column at 24.2 fb px
(lower band) / 24.5 (upper band) vs 24.9 for the right partial column -
full width, i.e. the window edge sits at 2H - 196 and exposes bitmap
sample 0. Copperline's 2H - 194 predicted 22 (2.8 photo px narrower than
measured, versus about 0.1 px scatter).

The old anchor also explains why the removed early-DDF "-1 sample"
correction existed: with the window edge 2px too far right, a standard DIW
overran the fetched row on early-DDF screens.

What changed

• bitplane::DIW_HSTART_FB0 and bus::RENDER_DIW_HSTART_FB0 0x61 -> 0x62
  (window comparator + sprites/collisions).
• Fetch references +1 (DIW_HSTART_FETCH_REFERENCE_LORES/HIRES, and the
  capture twins): bitmap sample positions are unchanged in the framebuffer.
• Derived values move automatically: BITPLANE_CONTROL_PIPELINE_FB 34 -> 36,
  STANDARD_VISIBLE_X0 64 -> 62, TV cutout, presentation recentring 26 -> 24.
• Colour/copper anchors (COLOR_WRITE_HPOS_FB0 etc.) untouched - they were
  verified against vAmiga (latch2 0.000%) and hardware.
• Test expectations updated with hardware-verdict comments; sprite clxdat
  tests keep their overlap geometry by moving the programmed sprite hstart
  with the comparator.

Scores and checks

• cargo test --release green (28 tests re-derived for the new anchor with
  hardware-verdict comments), cargo clippy and cargo fmt --check clean.
• Calibrated-web checks (fetch-reference gates): KS1.3 boot (hires FMODE=0)
  and A1200 KS3.1 boot (hires FMODE=3) differ from the previous stage ONLY
  in the 2px border column at the window edge - the hires pictures did not
  move. Second Nature (lores composite) and the KS1.3 TV-mode presentation
  eyeballed clean; TV recentring follows the window (26 -> 24).
• Demo screenshots vs the early-DDF stage: every difference is in the
  predicted classes - the newly exposed first bitmap column at standard
  window edges (Hamazing x 62-63), window-close clip at 702 (Inside the
  Machine), border column at x 14-15 (Gen-X, kicka1200, Zool), and sprite
  layers moving with the comparator (Zool cracktro text rows shift exactly
  -2 with zero residual; Gen-X gears shutter edge columns).
• cargo test --release --test image_regression -- --ignored: 7 pass
  (ocs_bpu7_ham is the documented host-flaky perf test).
• vAmigaTS full sweep (1929 cases vs vAmiga 4.4 refs), bucket means
  combined with the early-DDF PR against the pre-fix baseline:
  Agnus/DIW/OLDDIW 47.0% -> 4.3%, Denise/DIW 9.8% -> 0.9%, Denise/Sprites
  9.4% -> 5.1% (114 cases - the sprite anchor verified against vAmiga),
  Agnus/Copper 15.8% -> 9.7%, Agnus/DDF 9.7% -> 6.9%, Agnus/Blitter
  16.5% -> 14.4%. Standouts: Denise/Misc/Overscan/overscan2 90.8% ->
  0.004%, diw7/diw9/diw3/jumpbpu2 exactly 0.000%. The only score
  increases are CPU-timing buckets and coprace2; spot-checked MOVES8w
  (48.339%) and coprace2 (42.383%) regenerate to the same scores on
  clean main - pre-existing drift from the floating-bus work (vAmigaTS sweep fixes: 68000 floating bus + CIA-B POUT/CNT wiring #102),
  not this change.
• Gen-X issue Demo issues #69 "endpart leftmost column missing": the end-text pages
  now show the previously clipped first lores column at the window edge
  (620s screenshot, content at framebuffer x 62-63).