These should not have been public in the first place. This makes invalid
state between geometry and data impossible (fixing any problems related
to #61) and results in a nicer API overall.

Basically: Plane and AlignedData have safety invariants that depend on
the state of PlaneGeometry when `Plane::new` is called. By both
validating the geometry in its constructor and making it fully
immutable, we ensure that the invariants are always fulfilled.

This also means we can remove the re-validation of geometry (part of
#74) from `Plane::new`.

Changes:
- add getters
- make getters return zeroable scalars over NonZero<T>. With `#[inline]`
this should elide unnecessary `width != 0` checks downstream while
making the API simpler (`plane.width().get()` -> `plane.width()`)
- move data_origin from Plane to geometry, not something downstream
needs without padding awareness (-> `-F padding_api`)

#74 uncovered that empty data/len==0 returned a dangling pointer that
was not following the same alignment as real allocations. This required
some doc changes and is just not nice in general.

Use `layout.dangling_ptr` instead of the less specific
`NonNull::dangling_ptr` to follow our "usual" alignment rules exactly.
This requires an MSRV bump to 1.95.