Load deferred tool outputs in one request without scroll animation

[stippi]

Problem

Opening a session with several expanded tool cards whose outputs are stored as blobs replayed the "scroll to bottom" animation several times.

Cause: every expanded block with a deferred output requested it on its own, and each answer came back as UiEvent::UpdateToolStatus — the live agent path, which calls auto_scroll_if_following and starts the spring animation. scroll_to_end only holds until the next layout (it then becomes a regular top anchor), so each arriving output that grew a visible block pushed the bottom out of view again, and the animation chased it — once per output.

A second issue: those answers carried no session ID and were applied to whatever session was shown when they arrived, including triggering the animation there. Tool IDs are only unique per session.

Changes

• Core: SessionService::load_tool_outputs(session_id, tool_ids) replaces load_tool_output and reads all requested blobs in one call. A result that can't be read (or an unknown tool ID) is left out with a warning instead of failing the batch.
• New event: UiEvent::ToolOutputsLoaded { session_id, results } instead of reusing UpdateToolStatus. ACP ignores it.
• GPUI:
  • Blocks no longer dispatch the request themselves; BlockView::claim_deferred_output hands out the tool ID once, and the event loop collects all of them after applying tool results and sends a single request. Expanding an inline block sends a request for that one ID.
  • ToolOutputsLoaded for a session no longer shown is dropped.
  • On arrival, a list that follows its tail jumps to stay at the bottom (keep_tail_in_view) instead of animating; a list that doesn't follow keeps its anchor.
• Docs: docs/session-storage.md, AGENTS.md.

Tests

• loaded_outputs_keep_a_following_transcript_at_the_bottom_without_animating — failed on !animating before the fix (reproduces the bug).
• outputs_loaded_for_a_session_no_longer_shown_are_dropped
• Core test updated to the batch API, including an unknown ID being left out.

cargo test --workspace, cargo fmt --check and cargo clippy --all-targets --all-features -D warnings pass. Not yet verified in the running app.

Notes for review

• In-flight loads are not cancelled on a session switch; the blobs are read and the answer discarded. With one request per session open this stays bounded.
• In the expand path the block is marked as requested before checking for a session ID / Gpui global. Both always exist in the app; only relevant for tests without the global.