Allow images with different channel counts

Nobody has claimed this yet.

Currently, if you want to render an image, vello_cpu forces you do provide an RGBA image. It would be great if some custom color space layouts could be defined, in my case:
- RGBA
- RGB
- LUMA
- LUMAA

The advantage would be that the user doesn't have to reallocate a completely new buffer when converting to RGBA but can pass existing allocations, which saves a lot of memory. In my particular case, the motivation is that scanned black/white PDF pages are often very big, so in that particular case it's especially painful if you have to convert to RGBA. However, I'm also not sure it would be generally useful to others, and I see two concerns:
- If we implement this in the form of an additional generic parameter for the sampling, this could have bad effects on compile-time and binary size, especially since our current image code already relies a lot on generics.
- If we implement this in the form of a match statement, it could potentially reduce performance.

So I'm not sure if we actually should add this.

The issue names vello_cpu's image-rendering path but no files or tests. Start by tracing where RGBA input is required and compare representations for RGBA, RGB, LUMA, and LUMAA against the generic-versus-match tradeoffs described. Done means agreeing on an API and implementation approach that avoids unnecessary buffer conversion without unacceptable compile-time, binary-size, or performance costs.

Written by the indexing model from the issue text.