What problem does this feature solve?

Currently even if the module includes TLA, the modules are not executed concurrently.
It is described in https://rolldown.rs/guide/in-depth/tla-in-rolldown#concurrent-to-sequential.
But as noted in the document, this changes the semantics.

For example, this reproduction works before bundling, but stalls after bundling. Enabling exprimental.strictExecutionOrder does not help.

What does the proposed API look like?

Not an API change.