Add input to specify PNG, JPEG, or WebP output formats

[mattt]

Resolves #75

Currently, predictions generate images in PNG format. This PR adds a new format input that allows predictions to generate JPEG and WebP formats instead. If unspecified, the PNG is used by default.

[zeke]

So it automatically saves as each image type based on the extension of the target filename? 🤔

[mattt]

So it automatically saves as each image type based on the extension of the target filename? 🤔

Yes. This behavior comes from Pillow:

[T]he save() method looks at the name to determine which format to use, unless the format is given explicitly.

[zeke]

[mattt]

PNG and JPEG work correctly, but there's a bug in Cog around MIME types and data URI encoding/decoding that prevents WebP path outputs being saved to disk (at least when running `cog predict; I haven't tested running the model on Replicate).

Converting to draft until that's resolved.