Add reference docs, Cloudflare examples, and test runner

[andreasjansson]

• Split SKILL.md into focused reference docs: PREDICTIONS.md, MODEL_SEARCH.md, COLLECTIONS.md, WORKFLOWS.md, DEPLOYMENTS.md, HTTP_API.md, CLOUDFLARE_WORKERS.md, CLOUDFLARE_WORKFLOWS.md
• Add Cloudflare Workers examples: basic, async poll, concurrent, image editing via request.blob(), routing by path
• Add Cloudflare Workflows example: keyframe → video → stitch pipeline
• Vary models across examples (p-image, flux-2-klein, flux-schnell)
• Add script/test_snippets.py: extracts and runs all code snippets, spins up wrangler dev for worker/workflow snippets, detects image inputs and POSTs test fixture
• Add ruff check/format to script/lint
• Add test-snippets job to CI (skipped on fork PRs)
• Update AGENTS.md with new structure and testing docs

[zeke]

Pushed up some changes to prevent "bitrot" with hardcoded references to ever-changing models and collections.

But... blocking this pending further discussion, because:

I am unsure about all the client library usage.. there's a lot of usage in this PR that won't work with the new Stainless SDKs, and advertising these features in a skill would just create more debt for us later if we plan to ship those.

We don't really have a plan around what we're doing with our old and new Replicate client libraries in this new Cloudflare world.. my inclination is to let the alpha and beta Stainless SDKs die on the vine, because it won't be worth the disruption to users, especially if they're gonna have to migrate from Replicate API to Cloudflare API at some point anyway...

My inclination would be to keep the SDKs out of these skills entirely. Let the agents write code directly against the HTTP API. They won't get much benefit from the SDKs, and the API is way more stable and easy to reason about.

[zeke]

@andreasjansson as you mentioned in chat:

we can just update the skill and people’s agents will keep working (as opposed to the fixed-in-time knowledge the agent is pretrained with)

With that in mind, I'm okay with shipping this, but the changes as they are seem heavy-handed. I'd love to keep the skills lean so agents know everything they need to know to "self serve", instead of us feeding them a large and potentially overly prescriptive pile of docs full of subtleties that can lead them astray, like code snippets implicitly recommending models that will soon be outmoded.