This commit fixes a regression where `ngModel` no longer syncs
letter by letter on Android devices, and instead syncs at the
end of every word. This broke when we introduced buffering of
IME events so IMEs like Pinyin keyboards or Katakana keyboards
wouldn't display composition strings. Unfortunately, iOS devices
and Android devices have opposite event behavior. Whereas iOS
devices fire composition events for IME keyboards only, Android
fires composition events for Latin-language keyboards. For
this reason, languages like English don't work as expected on
Android if we always buffer. So to support both platforms,
composition string buffering will only be turned on by default
for non-Android devices.
However, we have also added a `COMPOSITION_BUFFER_MODE` token
to make this configurable by the application. In some cases, apps
might might still want to receive intermediate values. For example,
some inputs begin searching based on Latin letters before a
character selection is made.
As a provider, this is fairly flexible. If you want to turn
composition buffering off, simply provide the token at the top
level:
```ts
providers: [
{provide: COMPOSITION_BUFFER_MODE, useValue: false}
]
```
Or, if you want to change the mode based on locale or platform,
you can use a factory:
```ts
import {shouldUseBuffering} from 'my/lib';
....
providers: [
{provide: COMPOSITION_BUFFER_MODE, useFactory: shouldUseBuffering}
]
```
Closes#15079.
PR Close#15256
