[Core][API] get error from ObjectRef without the overhead of fetching the actual data

[scv119]

Description

user want to handle the error of ray tasks, without paying the cost of fetching the potentially large data if the task succeeded, there is no easy way to do so.

essentially we need an API like object_ref.on_error(lambda) where you only pay the cost of getting error message.

Use case

No response

[jovany-wang]

The pain point makes sense, but I guess it might break the async-post/get-future pattern right? although I know that some other frameworks do like this.

For examples, we should answer these questions first:

1. What's the thread that the lambda gets executed in?
2. What happens if the object_ref gets out of the scope?

[rkooo567]

I think we can follow the same API as concurrent.futures API here; I.e., supporting exceptions() API