fix(ime): implement zcr_extended_text_input_v1 handlers and fix serial/activation bugs

[kkimdev]

This is PR 2 of a stacked series (split from #27 for reviewability).

---

Main fix: last syllable lost when committing Korean text

Steps:

1. Switch to Korean IME input method
2. Type "가나다" (three syllables, Enter after each: ㄱ→가→가낙→가나→가나닥→가나다)
3. Press Enter to commit

Observed: "가나" appears in the application. "다" is lost.

Intended: "가나다" should appear in the application.

Root cause: When the ChromeOS IME sends on_commit_string followed by on_preedit_string(""), the final preedit clear was sent as a v3 preedit_string("") but with no following done event. The guest needs done after every state update to apply it; without it, the preedit clear is never processed, so the final commit appears alongside a stale preedit text which the guest may discard.

Fix: A done event is now sent after every preedit_string, commit_string, delete_surrounding_text, and set_preedit_region — not just after commit. A with_state helper provides a monotonically increasing done_serial for each such done event, replacing the hardcoded done(0).

confirm_preedit (Enter/Space to commit): Was a no-op. Now forwarded as v3 preedit_string("") + commit_string + done, so pressing Enter or Space to confirm preedit works.

---

Other changes necessary to make the core fix work without regression:

The fixes below are interdependent with the core fix — applying the done serial fix in isolation, for example, breaks on_keysym forwarding (causes replayed stale preedit text), and confirm_preedit needs activation + serial state to work correctly. They are included together to form a coherent, regression-free patch.

• on_set_preedit_region: Translates the extension event into v3 delete_surrounding_text + preedit_string + done, using cached surrounding_text from v3. Validates UTF-8 char boundaries. Was a no-op; ChromeOS IME calls this on every keystroke.

• update_host_activation (new): v3 enable/disable alone doesn't tell the host when to show the IME — the host also needs wl_keyboard.enter/leave. This state machine gates activation on (enabled && active_surface != None), called from keyboard.rs enter/leave and from v3 on_commit. Without it, applying the done serial fix causes the IME to activate on the wrong surface or not at all.

• store_host_serial + real serial in commit_state: v3 commit_state was sending serial 0, which the host skips. The serial from every v1 event is now stored and forwarded in commit_state. Required because the serial-aware done fix reveals that the host was also ignoring commit_state.

• surrounding_text_dirty flag: on_commit was consuming surrounding_text via .take() every time, so after forwarding once, subsequent calls sent nothing. Now it only forwards when the guest actually calls on_set_surrounding_text. Without this, set_preedit_region has no text to reference after the first edit.

• on_keysym forwarding: Forwards IME key events as wl_keyboard::key to the guest with the real host serial and time. Without this, the core done serial fix enters a stale-preedit-text regression because the host serial namespace doesn't align with the guest's expectations for key event ordering.

[InternetOfTofu]

Is there a sequence I should merge the 31-34 PR?

[kkimdev]

@InternetOfTofu PR 31-34 doesn't have inter-dependencies and can be merged in any order. There will be a PR(the main bug fix of this PR) that depends on prior PRs, but I will upload it after all the dependency PRs are merged.

I tested with the following command to see if there is an obvious regression for the individual PRs.

cargo build --release
(trap 'kill 0' INT; \
  cargo run --release -- wayland-test-1 & \
  sleep 5; \
  WAYLAND_DISPLAY=wayland-test-1 code)

I will convert this PR to a draft.

[kkimdev]

I can update PRs and resolve conflicts(if any) as you merge.