The old C-style `BLI_ASSERT_UNIT_V...` assert macros have a few issues:
* They are named `unit`, but also consider a zero-length vector as valid.
* They use a fairly high epsilon value, which was defined because
vertex normals used to be stored as shorts.
Fortunately, these are used only in one place in the modern BLI_math C++
code AFAICS, which is `math::rotate_direction_around_axis`.
This commit adds some utils to check for vectors being (almost) unit
or zero length, using more modern bases for epsilon values (from
`std::numeric_limits`).
* `is_zero` keeps its existing default arror of `0` (i.e. strictly null
vector by default). That way, current behavior is not changed, and in
most cases null vectors are explicitely created as exactly null.
* `is_unit` uses a default 10 times the type's epsilon, as a zero
epsilon would virtually never succeed here.
And it modifies `rotate_direction_around_axis` to:
* Assert that `axis` is a unit vector.
* Early-out in case given `direction` is a null vector, or rotating
angle is zero.
* Assert about `direction` being a unit vector otherwise.
Note that this will make `rotate_direction_around_axis` use much
stricter epsilon error factors. This does not seem to affect any of the
files that triggered asserts prior to recent fix in e18dd894b8 though.
Pull Request: https://projects.blender.org/blender/blender/pulls/122482
dd3222d2f8