security: fix SSRF in load_image and load_video URL fetching

[hoangperry]

Summary

load_image() and load_video() call requests.get(url) with no URL validation. When diffusers is deployed as a server-side inference API, any user who supplies image/video URLs can cause the server to make HTTP requests to internal network addresses including cloud metadata endpoints.

Vulnerable code (before)

# load_image
image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)

# load_video
response = requests.get(video, stream=True)

requests.get follows HTTP redirects by default, so an attacker-controlled server returning a 302 to an internal address bypasses any future URL-based filtering.

Fix

Add _validate_url() that resolves the hostname to IP and blocks RFC 1918 (10.x, 172.16.x, 192.168.x), loopback (127.x), and link-local (169.254.x) ranges before any HTTP request is made.

Test plan

• http://127.0.0.1/image.jpg → ValueError (loopback)
• http://169.254.169.254/image.jpg → ValueError (cloud metadata)
• http://10.0.0.1/image.jpg → ValueError (RFC 1918)
• https://upload.wikimedia.org/image.jpg → works normally