### TL;DR
Refactor space-triggered AI Widget activation logic from `keydown` to `keypress` event listeners
### Background
The `keydown` event triggered by a space may originate from:
1. Normal space insertion
2. Space triggered by input method confirming candidate words
In scenarios like (2), some browsers (see [ISSUE](https://github.com/toeverything/AFFiNE/issues/11541)) and input method callbacks produce events identical to scenario (1),making it impossible to distinguish between the two.
To fix this, the space-activated AI listener uses the `keypress` event:
In scenario 2, `event.which !== 32` (may be `30430` or other values) can be used to differentiate from scenario 1.
> CLOSE BS-3081
Close [BS-2866](https://linear.app/affine-design/issue/BS-2866/presentation-mode中的note消失)
## Problem
When using RequestAnimationFrame (RAF) for GFX block updates, there was a timing issue where the transform update would lag behind the RAF callback, causing the block to display with the previous frame's transform state.
## Solution
1. Refactored the block state management to use signals for better reactivity
2. Moved block visibility state management from `viewport-element.ts` to `gfx-block-component.ts`
3. Added `transformState$` signal to track block state
4. Synchronized transform updates with RAF using `effect` to ensure updates happen in the correct frame
5. Added test case to verify note visibility in presentation mode
In this way, all downstream callers can be guaranteed by correct viewport fit result, instead of requiring them to set `forceUpdate: true` param explicitly to them. The resizing optimization is an internal exception.
