Custom models: forward PDFs to the provider instead of extracting text

Built-in models receive attached PDFs directly and can analyse visual content

such as charts and diagrams. Custom models receive only client-side extracted

text, so a PDF whose content is image-based or vector-drawn arrives with

nothing usable, or with nothing at all.

Verified with the same file, same device and same session: built-in Sonnet 5

read a vector chart correctly, while a custom OpenRouter Sonnet 5 pointed at

the same underlying model returned "Could not read PDF text content." Support

confirmed this is current behaviour rather than a bug.

Request: let custom models receive the raw PDF, ideally via a capability

toggle alongside the existing "Support Image Input" so it is only sent to

endpoints that can handle it.

This matters for anyone routing through a custom endpoint (proxy, gateway or

self-hosted worker), where the built-in-model workaround is not available.