Faster stable diffusion

[emilk]

The stable_diffusion example had a bug that caused two parameters to be hard-coded (guidance_scale and num_inference_steps, even though there were arguments to ostensibly set these.

num_inference_steps was hard-coded to 70, which means the example takes a looooong time to run on the CPU. This in-particular would grind just py-run-all to a half during testing.

I fixed the bugs, simplified the code, and lowered num_inference_steps to 10. The result is quite noisy, but serviceable, and users can (now) control it with --num_inference_steps.

Checklist

• I have read and agree to Contributor Guide and the Code of Conduct
• I've included a screenshot or gif (if applicable)
• I've added a line to CHANGELOG.md (if this is a big enough change to warrant it)