Consider optimizing `LeftSemi` to emit output incrementally

### Is your feature request related to a problem or challenge?

Right now, `LeftSemi` consumes the entire probe side before emitting _any_ output rows. This isn't fundamental: we could check which probe row flips the visited bitmap from 0 -> 1, and arrange for the output row to be produced at that point. This would enable better pipeline parallelism.

It would probably also improve partitioning behavior: `LeftSemi` + `CollectLeft` emits everything in a single output partition today, which isn't great.

Would need to check if this needs stronger memory ordering requirements when flipping bits than we can get away with today.

### Describe the solution you'd like

_No response_

### Describe alternatives you've considered

_No response_

### Additional context

_No response_