Replace linearized polyline sampling with direct cubic Bézier evaluation
when the input is a BezierPath. This gives smooth, continuous tangent
rotations instead of the piecewise-constant tangents from the 30-point
linearization.

Key additions:
- Cubic Bézier eval/deriv/second_deriv functions
- 8-point Gauss-Legendre arc-length quadrature
- Offset-curve arc length via curvature formula (speed·|1-d·κ|)
- Binary-search inverse arc-length for parameter recovery
- Control-point extraction/reassembly for projecting BezierPath to pixel
- Separate convert_arguments methods (no more PointBased trait)

Layout RichText using layout_text() to get per-glyph positions, colors,
fonts, and sizes. Each glyph is wrapped as a single-char RichText
carrying its own style, letting the child text! handle font/color/size
natively per block.

Example: pathtext!(ax, path; text = rich("A", rich("B"; color=:red, font=:bold)))

The glyph y-origins from layout_text (e.g. subscript below baseline,
superscript above) are now applied as perpendicular shifts from the
path at each glyph position.

layout_text returns absolute y-origins (baseline at ~ascender height,
not y=0). Subtract the baseline y so regular glyphs sit on the path
and sub/superscripts shift perpendicular relative to it.

…tion

align now accepts a (halign, valign) tuple like the text recipe:
- halign: :left, :center, :right, or Real fraction (along path)
- valign: :baseline, :bottom, :center, :top (perpendicular to path)

valign uses font ascender/descender metrics to shift the text
perpendicular to the path. offset remains as an additional manual
pixel shift on top of valign.

Each glyph's rotation is now computed from the chord between its start
and end positions on the path (spaced by the glyph's advance width),
rather than the instantaneous tangent at the origin. This gives wider
characters like "W" a rotation that better matches the arc they span.

- docs/src/reference/plots/pathtext.md with BezierPath, polyline,
  and RichText sub/superscript examples
- attribute_examples for text (String vs RichText), align (valign
  comparison), and offset

Five reference tests covering:
- align: halign (:left/:center/:right) and valign (:top/:center/:bottom)
- String with per-char color vector and RichText with sub/superscripts
- offset: perpendicular shift from path
- polyline with NaN sub-path gap, BezierPath with LineTo/CurveTo/EllipticalArc
- space = :data vs :pixel for the path coordinate system

When a glyph's start and end samples land on different sub-paths
(separated by a NaN in a polyline or a MoveTo in a BezierPath), use
the tangent at the start position instead of the chord, which would
otherwise connect unrelated points across the gap.

- Shrink all pathtext refimages; hide axis decorations for denser framing.
- Combine "string+richtext" and "offset" tests (both shown together on
  one curve with opposite offsets and per-char vs RichText styling).
- Shorten polyline and BezierPath demo text to fit their smaller axes.

…ation

Split the text and path sides via inner-method dispatch:
  _layout_glyphs(text, ...) — AbstractString vs RichText
  _prepare_path_sampler(path, ...) — polyline vs BezierPath
The outer _pathtext_layout is now one method that orchestrates both.

…t-length

- _sample_bezierpath_at now binary-searches a cumulative arc-length
  table precomputed in _prepare_bezierpath, instead of scanning segments
  linearly per sample.
- _cubic_inv_arclen uses 20 bisection iterations instead of 30
  (≈1e-6 parameter precision, well below sub-pixel).
- _layout_glyphs computes x_positions with a single-pass accumulator
  instead of cumsum(advances) .- advances.
- total_text_len is now computed inside _place_glyphs_on_path instead
  of passed in (one fewer positional arg).
- Trimmed a few WHAT-comments; removed an unreachable return.