Fix bug when fatal error is thrown as a result of batch.commit

[acdlite]

Fixes #12474

[gaearon]

Seems okay.

Can you give me a rough explanation of the top-level flow these days? I'm not sure how nextFlushedRoot, nextFlushedExpirationTime, nextRoot, nextRenderExpirationTime, and root's expirationTime interact now.

[acdlite]

It's a bit awkward because we're thinking we might lift the root scheduling stuff at the bottom (performWork and all that) into the renderers. Or at least one level higher. That way renderers like React Native have the option of using a different root scheduler than the one used by React DOM.

That's why you see this duplication of nextFlushed<Thing> vs next<Thing>. It's an artifact of the separation between the renderer and scheduler. The ones with flush in the name conceptually belong to the renderer.

General flow is

• setState fires requestWork, which adds a root to the root schedule and schedules a callback (performAsyncWork) with requestIdleCallback.
• performAsyncWork calls performWork, which loops through the list of roots and starts working on them in order of priority.
• performWorkOnRoot calls the render phase and commit phase as separate entry points. This way, a renderer has the option to defer the commit phase until later. That's how createBatch works. Before, the commit phase would always fire right after the render phase, so there was less control.