Make the behavior consistent for asyncio and asynq: raised error.

[SoulTch]

This PR addressed two behaviors:

• Propagates raised errors to the caller so that the caller can catch the exception.
• If an asynq function yields a sequence of tasks, cancels all other tasks if any of them fail.

[dkang-quora]

asynq does not cancel other concurrent tasks.
If you gather two independent tasks like two memcached requests to two hosts, a failure from one should not cancel the other one.

import asyncio

from asynq import asynq

counter = 0

@asynq()
def f():
    global counter
    counter += 1
    if counter == 3:
        raise RuntimeError('error 3')


@asynq()
def g():
    print('<<')
    try:
        yield [f.asynq(), f.asynq(), f.asynq(), f.asynq(), f.asynq()]
    except Exception as e:
        print('except', e)
    print('>>')


g()
print('counter=',counter)

<<
except error 3
>>
counter= 5

[dkang-quora]

Can you fix this with return_exceptions=True? See asyncio.gather(*aws, return_exceptions=False)

[dkang-quora]

Wait, no. We want return_exceptions=False. The current code does not cancel on the first failure.

[SoulTch]

Oh, I naively thought asynq will cancel the remaining tasks. Thanks for checking this.
So, which behavior is preferred?

• If any of task raised an error, return immediately. Other tasks will remain in the event loop.
• Wait all tasks. If any of them raised an error, raise any (or we can choose which one to raise).

It seems second one is more close to existing asynq.

[SoulTch]

Changed to:

• Wait all tasks even though some of them raised an error. If two or more errors were raised, raise any of them.