The extraction overlay currently only allows expanding/collapsing step
output after the entire pipeline finishes. If the LLM step takes 300s+
(e.g. qwen3:0.6b on OCR text), the completed text and OCR results sit
there unreachable behind a spinner.

Expected: each step's output is immediately expandable as soon as
that step completes, even while later steps are still running. The
cursor and expand/collapse keybindings should work on completed steps
at any time.

This also means the "accept" action could partially apply — persist
text/OCR results without waiting for LLM to finish, or let the user
cancel just the LLM step while keeping the rest.

Motivation: observed qwen3:0.6b taking 300s+ on a single image.
The OCR text was ready in <5s but couldn't be inspected until the
LLM step (eventually) finished or was cancelled.