Unify background handling under a single :background option

[hlindset]

First of all: Very open for input/suggestions/discussions on the choices made in this PR!

In an earlier iteration of the PR, I removed libvips' default fill (the zero-fill) and instead leaned on using background: :transparent + extend_mode: :background, but that caused some inconsistency with e.g. CMYK on transform ops vs flatten/write, so I reversed course and instead standardized on default/background: nil meaning libvips' default zero-fill, which resolved the inconsistency. The only wrinkle is that libvips' default is only accessible by not setting :background (or explicitly setting it to nil).

---

Generally

• Collapse the :background_color/:background_transparency options into a single :background option, resolved via Image.BackgroundColor.resolve/2 everywhere (embed, flatten, join, rotate, affine, translate, shear, warp_perspective, write)
• Transparency can be set/overridden by wrapping any color (including :average) in a tuple, like so: {color, alpha: a}, where a is 0..255 | 0.0..1.0 | :opaque | :transparent (the Image.Pixel.transparency type)
• background: nil means "unset" everywhere (falls back to the op's default).

Defaults

• An omitted (or nil) :background defers to libvips' native all-zeros fill everywhere (embed, join, rotate, affine, translate, shear, flatten, write): transparent on images with an alpha band, black otherwise (white in CMYK). Keeps flatten() |> write() and write() consistent.
• Exception: warp_perspective defaults to :black (it flattens before warping)

extend_mode

embed

• Reduced to :copy | :repeat | :mirror | :background
• :black/:white removed: use an explicit :background instead. They're somewhat confusing anyway, e.g. :black actually being white in CMYK.
• A content mode (:copy/:repeat/:mirror) consumes no color, so combining one with a non-nil :background is contradictory and returns an error

affine/shear/translate/warp_perspective

Extend here only affects the antialiased fringe at the content edge. The uncovered canvas is always filled with :background. extend_mode: :background is the option that makes sense most of the time.

• Reduced to :background | :copy
  • extend_mode: :background (default) blends the fringe into the fill, which is right whenever canvas is exposed
  • :copy clamps to content, which can be right e.g. when the content fills the whole canvas, where using extend_mode: :background would leave a faint border along the edge
• Removed :repeat/:mirror modes
  • :repeat bleeds the opposite edge into the fringe
  • :mirror is pixel-identical to :copy for the fringe
• rotate doesn't accept :extend_mode. Both of its internal paths blend the fringe toward :background.

Breaking changes

• :background_color / :background_transparency → :background (embed, flatten, join)
• embed on an alpha image now defaults to a transparent border (was opaque black)
• embed's :black/:white extend modes removed
• affine-family :extend_mode narrowed to :background | :copy
• flatten with no :background now defers to libvips' native fill instead of :black: identical for sRGB, paper-white instead of ink-black for CMYK (aligning flatten() |> write() with write())
• embed with content :extend_mode (:copy/:repeat/:mirror) plus a non-nil :background now errors

P.S. warp_perspective/straighten_perspective currently flatten. I can have a look at solving it similarly to affine with premultiply/unpremultiply to support transparency through the transform in a follow-up PR, resolving the one exception left after this PR