Fix hang up when loading duplicated CSS module chunks (#4879)
### Description
I identified this issue when authoring a test for PostCSS. Here's how it
goes:
1. CSS chunk 1 loads with module chunk A and source = runtime. This
marks chunk A as an available module chunk. The promise for A is the
same promise as for CSS chunk 1, which is resolved in registerChunk
because chunk 1 is part of the chunk group of the initial JS chunk.
2. CSS chunk 2 loads with module chunks A and B. Since A is already
available, we run into a different branch, where we only try to load B.
However, since we're still using source = runtime, we don't actually
attempt to load B and instead return a promise to the resolver of B. But
B isn't part of the chunk group of the initial JS chunk, so it is never
resolved.
This hangs up the runtime entirely and breaks page hydration.
Instead, we move the "automatically resolving CSS chunks when source =
Runtime" from `registerChunk` (which wasn't quite correct either as far
as I can tell) into `loadChunk`, which solves this issue.
### Testing Instructions
I don't know if it's worth having a test case just for this, but this
will technically be covered under the PostCSS test case in
https://github.com/vercel/next.js/pull/49463.
link WEB-1023
