Use enlargeable types in to_luma functions.

[197g]

This is an attempt to fix #1214, using the Enlargeable trait to
select a suitable type for the calculation in rgb_to_luma and
bgr_to_luma.

Integer pixel types will use suitable integer pixel types for the
calculation, while floating point pixel types will use floating point
types.

This PR also provides the Enlargeable trait for all Primitive
types (luckily, Primitive is not implemented for 128 bit numeric
types).

The Color implementation of all library-defined structs is then
restricted to a subset of Primitive channel types. This is necessary to
allow its conversion operations to utilize the private Enlargeable.
Notably however we relax the type itself to not utilize any trait
bound. It's a simple wrapper around an array now.

Rebase of #1215 which was auto-closed due to deleted base branch..

[197g]

@fintelia Other concerns with the public change of this PR? It overlaps with reworking the Pixel trait in that if we removed invert and color conversion as originally sketched in #1099 then it might be possible to relax this to Primitive again.

[197g]

These are the two API breaking issues still open from 0.24. We could also do a release just to get it out and then work hard on actually fixing color, with conversion and better buffer API, in a 0.25.

[fintelia]

I think maybe merge this and then roll out a 0.24? A broader Pixel rework for 0.25 seems reasonable to me.