Return ContentThinking objects from stream_text()

[simonpcouch]

Addresses #837, and a subset of #828. Companion to posit-dev/shinychat#167.

Changes stream_text() to return Content objects instead of strings, allowing downstream consumers to distinguish between content types during streaming. With the companion shinychat PR, results look like this:

ellmer-shinychat-thinking.mp4

I'm not sure if this is the opposite direction of the dev dependency that yall would usually introduce. With main shinychat and this PR, each thinking delta would otherwise get its own collapsible:

So, I opted to make ellmer depend on the shinychat Remotes. If this is a no-go, we could add some version-dependent logic, but felt cleaner to start with this approach.

Here's the reprex I use in the video:

# pak::pak("posit-dev/shinychat/pkg-r#167")
library(ellmer)
library(shinychat)

# Anthropic with thinking ------------------------------------------
chat_anthropic <- chat_claude(
  model = "claude-sonnet-4-5-20250929",
  params = params(reasoning_tokens = 1024)
)

clipr::write_clip(
  "Write a limerick about the tidyverse, returning only the limerick."
)
chat_app(chat_anthropic)

# OpenAI with reasoning ------------------------------------------
chat_openai <- chat_openai(
  model = "gpt-5.1",
  params = params(reasoning_effort = "high")
)

chat_app(chat_openai)

# Gemini with thinking ------------------------------------------
chat_gemini <- chat_google_gemini(
  model = "gemini-3-flash-preview",
  params = params(reasoning_tokens = 1024)
)

chat_app(chat_gemini)

[simonpcouch]

I'm not sure if using switch on S7 objects is considered bad practice, but the control flow was gnarly otherwise😅

[hadley]

Maybe we should have a stream_data() that returns the object and then stream_text() would contain this code to extract out just the text?

[hadley]

Or stream_content()?

[hadley]

This feels inside out to me? Why doesn't stream_text() call stream_content()?