GeezerVison

`domain` is a plot fraction, so the gap it leaves between subplots grows and
shrinks with the figure while a subplot title does not. domainpad reserves a
fixed number of pixels inside the domain edges instead, resolved at draw time.

Applied in ax.setScale to _offset and _length rather than to ax.domain, so the
domain keeps meaning what the user typed and anything referenced to
`<axis> domain` follows the padded area without new drawing code.

Along with it:

- scattergl, splom and the rangeslider rebuilt the plot rect from `domain` and
  so did not see the pad; they now read _offset and _length
- pads that do not fit are backed off together, since setScale already throws
  on a negative length
- updateDomain scales the padded length rather than the raw span, so a
  scaleanchor ratio stays exact when the constraint shrinks a padded axis

Refs plotly#7835

Removed attribution from the commit message.

The three places that build the plot rect from `domain` were changed to read
_offset and _length instead. That round trips through pixel space, and the
result is not exactly the value it started from: splom cell domains came back
about one ulp out, which moved rasterised markers enough to fail the
splom_*-nodiag image baselines even though no padding was involved.

Add the resolved padding to the original expressions instead, and expose it as
ax._padStart / ax._padEnd for those callers. Adding a literal zero is exact, so
an unpadded plot now produces the same numbers it always did.

Also list domainpad in swapAxisGroup's noSwapAttrs, next to domain: its keys are
named for screen edges, so swapping x and y would need them remapped.

updateDomain multiplied the domain by (newSpan / inputSpan). Without padding,
that ratio is 1 / factor, but the multiply can change the last bit of the
result, so a plot with no domainpad could move by a tiny amount.

Divide by `factor` as main does, and change the divisor only when the axis
has padding. Without domainpad, the result is now the same as on main.

… and types

The comments and attribute descriptions now follow the Simplified Technical
English rules in AGENTS.md. Lines that this PR touches now use const, let and
arrow functions. There is no change in behavior.

The regenerated schema.d.ts and plot-schema.json fix the
generated-types-drift check, which failed because the types did not include
domainpad.

Spikes with spikemode 'across' got their span from the counter axis domains,
so with domainpad they crossed the reserved space, for example a subplot
title. Get the span from the domain minus the padding of each counter axis.

Without domainpad the span is the same as before. A free axis still extends
the span to its position. The new test fails without this fix.

scattergl and splom build their regl viewports from `domain`, so they add the
padding in their own code. These tests make sure that the viewport, the splom
cells and the regl grid match the padded plot area.

The mock shows four subplots with different domainpad sides, so a change in
the padded plot area shows in the image tests. The baseline image must come
from the baselines-default-diff CI artifact.

An earlier commit removed the "with thanks to" credit by mistake. The draftlog
convention includes it.

# Conflicts:
#	test/jasmine/tests/hover_spikeline_test.js

KoolADE85 left a comment

Remove the inline comments flagged in review, and describe in the
domainpad attribute that an overlaying axis uses the domainpad of the
axis it overlays. Regenerate the plot schema and the generated types.

Give the overlaying axis in the domainpad test a conflicting domainpad, so the test covers the behavior that the attribute description now documents.