Align colors in SDL package to Go specification

[mokiat]

Closes #580

This PR does the following changes:

• It ensures that Surface.Set does not panic by handling colors correctly.
• It changes Surface.At to return color.NRGBA instead of color.RGBA which is the correct type for what is being returned (non-alpha-premultiplied color). This can be a breaking change if anyone ever expected a specific color from the method, despite the interface abstraction.
• It changes the sdl.Color type to color.NRGBA which is the correct representation of what is held in sdl.Color (non-alpha-premultiplied color). This is a breaking change. Preserving the color.RGBA type mapping would require more code rework.
• It fixes all of the color models to properly return 16bit-component alpha-premultiplied colors.
• It fixes a broken type reference in the sdl package

Things to note:

• The color.RGBA type in Go represents an alpha-premultiplied color and SDL2 appears to use standard non-alpha premultiplied colors everywhere. This is why color.NRGBA is the correct pick.
• The Color.RGBA method is expected to return 16bit (i.e. from 0x0000 to 0xFFFF) alpha-premultiplied color components.
• The PR makes heavy use of color.NRGBA and color.NRGBAModel, since these are the official ones and they have stable and optimized implementations for color conversions between alpha-premultiplied and non-alpha-premultiplied variants and from 8bit-component to 16bit-component colors.

Side notes:

• In the past I had written a blog post about Go's colors which can shed some light on the above considerations. Feel free to remove this section if such type of links are unacceptable here.
• While downscaling of bytes is trivial, upscaling is not so if you want to fill the whole scale. I have tried to come up with the proper bit-wise operation expressions but to double-check I have used the following Go program to empirically measure it: https://play.golang.com/p/fYrvt4TXFv2

[mokiat]

This PR should be ready for a more general public verification. The more people that have color scenarios in their apps to verify the better.

[Lundis]

Great job!

I ran into the same issue today. This PR solves all of my issues and does not seem to break anything. NRGBA should be correct indeed 👍

[veeableful]

Thank you very much @mokiat! I will merge this into the v0.4.x branch and create a new patch release as well so more people can test it.

[mokiat]

Thanks! Using v0.4.36, colors work well for me now.

On a side note, asking for future reference: should I have opened the PR against v0.4.x or was master ok, since there seem to be other API changes in master (e.g. non-ptr events) that have not gone into v0.4.x?

[veeableful]

Yes, master is preferred in most cases. I usually cherry-pick the commit into v.4.x branch afterwards. (Speaking of, I really should get v0.5.x out already 😓, just stuck on the static libraries currently).