I'm having a little trouble with the performance of the library - my use-case is more-or-less exactly as [_example/mjpegproxy.go] - I'm reading several MJPEG streams and re-serving them. Looking at the code I notice that it effectively reads the JPEG bytes from each part, decodes that into an image.Image, then turns that image.Image back into bytes, which seems very inefficient?

I've tried skipping the JPEG decode -> encode and just writing the bytes from the part straight to the stream, but the output comes out very broken. I'd love to find a way to achieve this without needing the decode/encode step as I'm finding that's putting a lot of pressure on the CPU with multiple streams - I'm seeing it take on average between 50ms and 80ms per part, which isn't great for a 25fps stream.

I'll open a PR if I figure out a way to handle this, but otherwise I'd love to know what the decode->encode is doing under the hood, because it's clearly necessary!

Seems I can use this package to decode Motion-JPEG from Raspberry Pi V4L2, but what value should I send to second argument boundary? Will be good to have it in func docs. Thanks!

I don't see a license here, which means I can't use this for commercial work. If it's alright with you, I'd prefer MIT, Apache-2.0, BSD or some other liberal license.