Use only specific lyon packages instead of pulling all of them in.
This slightly speeds up compilation as well as for example lyon_tesselation
doesn't need to be compiled anymore.
Allow converting a brush to a color. In the case of a gradient, the color of the first stop is returned.
For the C++ generator this requires adding the extra case of explicitly
calling the `Brush(const Color &)` constructor, despite it being implicit,
in order to generate the correct code when we have IR that casts twice:
```
Expression::Cast {
from: Expression::Cast {
from: Expression::Cast {
from: Expression::NumberLiteral(...),
to: Type::Color,
}
to: Type::Brush,
},
to: Type::Color,
}
```
* Always apply a transformation (less variants)
* Let the path data iterator take ownership of the data. That will
make it possible to return a PathDataIterator from a function in the
future
We're going to need support for InterpolatedValue, which at the moment
has trait requirements that are a bit too tight.
It's sufficient if it takes references to values and doesn't require Copy or Clone.
The latter can go into the bounds where it's actually used.
A few changes were required:
* `LinearGradient(LinearGradient)` as enum variant unfortunately
won't compile because the cbindgen generated constructor
function (`LinearGradient()`) will try to also instantiate the
variant type inside (`LinearGradient`) and that won't find the type
but the function itself and error out. So the inner type is now
called `LinearGradientBrush`.
* The same name dance was required for `Color`, where the enum variant
instead is called `SolidColor`
* `BrushInner` was removed in favor of just `Brush`. The nicer Rust
API will be the public variant, and for cbindgen we can just put
the generated enum into an internal namespace, like we do for
Resource for example
* A `NoBrush` variant was added. Maybe that name could be improved?
