docs: add LLM equivalents for Call Flow screenshots

[Devon-White]

Summary

• hide Call Flow Builder canvas screenshots from Markdown exports
• preserve node topology, output-connector routing, variables, and configuration as text
• cover procedural builder screenshots and provider-specific TTS selections
• add local/manual deployed-preview assertions for all 22 affected routes and complex branch behavior; the checker is intentionally not wired into CI

Scope

This is PR 5 in the stacked implementation for #388 and depends on #613.

Validation

• yarn test:scripts
• yarn fern-md-check
• git diff --check
• yarn check-llm-exports passed for all 122 manifest routes
• source audit confirms every rendered Call Flow image is inside llms-ignore

Refs #388

[github-actions]

🌿 Preview

https://signalwire-preview-docs-614.docs.buildwithfern.com/docs

---

View workflow run

[hey-august]

Approving.

• All rendered Call Flow images verified inside <llms-ignore> (re-checked post-rebase); every hidden screenshot has a text equivalent or covering prose.
• Compared all 29 screenshots against their equivalents — node names, connectors, branch routing, and configured values match, including the corrected Gather Input loop-protection branches.
• All 22 manifest routes match page slugs.

Two nits, fine to fix here or in a follow-up:

• forward_to_phone.mdx: "Success completes the forward" — in the screenshot, Success routes to a Play Audio or TTS node; only the four failure outputs go to Hang Up Call.
• TTS/google.mdx / TTS/polly.mdx CFB tabs: gcloud.en-US-Neural2-A / amazon.Ruth:neural are SWML voice IDs, but the dropdown shows en-US-Neural2-A (premium) / Ruth (premium).