When the user pastes or drags-and-drops an image into a terminal that has
NLD on with autodetection enabled, entering agent view should lock the
input to AI mode for the duration of the attach. Previously, two leak
sites would unlock the input and let the classifier flip the buffer back
to shell mode after the entry:

1. The `EnteredAgentView` subscriber in `BlocklistAIInputModel` would set
   `is_locked: !is_ai_autodetection_enabled` even for `ImageAdded`
   origins.
2. `set_input_mode_agent`'s `should_unlock` branch (which is intended for
   the `!`-shell-toggle clear path in fullscreen agent view with empty
   buffer) would re-unlock immediately after the image-add path entered
   agent view.

This change introduces a single `has_locking_attachment()` predicate on
`BlocklistAIContextModel` and uses it to gate both leak sites. It also
adds an image-attachment-in-progress counter so the predicate is true
during the async file-read + resize/encode pipeline, before the final
`ImageContext` is appended to `pending_attachments`.

Changes:
- Add `pending_image_attachments_in_progress: usize` and
  `has_locking_attachment()` / `note_image_attachment_started()` /
  `note_image_attachment_completed()` to `BlocklistAIContextModel`.
- Wire a handle to `BlocklistAIContextModel` into `BlocklistAIInputModel`
  via a `set_ai_context_model` setter, called from `TerminalView::new`
  after both models exist.
- Add an `origin_requires_locked_ai` helper and force-lock to AI in the
  `EnteredAgentView` subscriber when the origin is `ImageAdded` or when
  `has_locking_attachment()` is true.
- Defense in depth: gate
  `BlocklistAIInputModel::is_autodetection_enabled_for_current_context`
  on `!has_locking_attachment()`.
- Increment / decrement the in-progress counter synchronously in
  `read_and_process_images_async` and `process_and_attach_images_as_ai_context`
  so callers see `has_locking_attachment() == true` for the entire
  attach pipeline.
- Reorder `handle_pasted_or_dragdropped_image_filepaths` and
  `process_and_attach_clipboard_image` so the editor's image dispatch
  runs before `set_input_mode_agent`, ensuring the counter is already
  incremented when the lock decision is made.
- Gate `set_input_mode_agent`'s `should_unlock` on
  `!has_locking_attachment()`.

Co-Authored-By: Oz <oz-agent@warp.dev>

Cover `has_locking_attachment` (default false; true on pending block ids,
selected text, or pending image; ignores file-only attachments; true while
the in-progress image-attach counter is non-zero) and the
`note_image_attachment_started` / `note_image_attachment_completed` counter
mechanics including saturating decrement at zero and multiple concurrent
in-flight pipelines.

To keep the test fixture small, add `#[cfg(test)]` `new_for_test`
constructors on `BlocklistAIContextModel` and `AmbientAgentViewModel` that
skip subscriptions/singleton lookups (`BlocklistAIHistoryModel`,
`LLMPreferences`, `CloudModel`, `UpdateManager`, `AppExecutionMode`). This
follows the existing `new_for_test` convention used by `Sessions`,
`TerminalModel`, `AuthManager`, `BlocklistAIHistoryModel`, etc.

Tests live in `app/src/ai/blocklist/context_model_test.rs` per the repo's
`${filename}_tests.rs` / `mod_test.rs` convention.

Co-Authored-By: Oz <oz-agent@warp.dev>

Switch from ctx.spawn to ctx.spawn_abortable in process_and_attach_images_as_ai_context so pending_image_attachments_in_progress is decremented on every exit path:

- on_resolve (normal completion): unchanged

- on_resolve early-return when handle was taken between value production and dispatch: now decrements before returning instead of leaking

- on_abort (future aborted before producing a value, e.g. submit_ai_query calling abort_attached_images_future_handle while processing was in flight): new explicit handler that decrements

Previously ctx.spawn's default no-op on_abort callback meant that aborting the spawned future left pending_image_attachments_in_progress nonzero, keeping has_locking_attachment() true and force-locking the input into AI mode until an unrelated context reset.

Co-Authored-By: Oz <oz-agent@warp.dev>