Don't memoize results that had an error

[megawac]

fixes #1465

This one is debatable if we want to change behaviour in 2.x. I think it makes sense to however could be swayed otherwise. Thoughts @aearly?

[hargasinski]

We should keep this as is for now, but for v3, I'd be open to changing this behaviour, as there's a potential race condition here (similar to #1248).

If the initial call to the memoized function takes between m and n seconds to err, and a second call is made x seconds after the initial (m < x < n) then the function might only be run once depending on whether or not the initial call finished. To eliminate it, we would have to call the memoized function again if there are queued calls, and an error occurred.

[aearly]

Yes, this is a potential 3.0 change.

It's debate-able as to whether caching errors is "correct" or not, though. e.g. caching the result of a request that 404's -- do you cache the 404, or keep retrying?

With this change, you could use reflect and always cache if you wanted, or use it normally and not cache errors.

[aearly]

This is ready to be merged for 3.0. Also, this new behavior is much less surprising if you were using an async function.