Move `.recipient()` from `Addr` to `Recipient`, Capacity fields public

[strohel]

Really make the errors() test deterministic

I claimed this in the previous PR, but it wasn't true: if the actor spawned quick enough the test would fail. What was true is that I've ran the testsuite hundreds of times in a loop, and it didn't fail. Some strange coincidence?

Make it truly deterministic by not spawning the actor at all (test using a naked Addr).

Move .recipient() from Addr to Recipient

This makes the API more flexible as we now can call .recipient() multiple times in a row on the same value, allowing one more M -> N message conversion in each step.

Previously, we thought it is too wild to create "arbitrarily long conversion chain", so we've put .recipient() on Addr, only allowing one such conversion.

But with the advent of actor wrappers like Timed<A>, this first conversion is "consumed" by the wrapper itself, limiting their users.

The method name Recipient::recipient() seems a bit weird, but there's a high value in keeping the current usage working without any changes (because Addr derefs to Recipient), and no obviously better name came into my mind.

Make Capacity fields (high, normal) public

This is an API omission discovered when we tried to set both normal and priority inbox sizes to different and non-default values.

I argue that Capacity { normal: Some(3), high: Some(9) } is more explicit than Capacity::new(3, 9), hence making the fields pub instead of adding a constructor.

---

Sorry for cramming this into one PR (the commits are still atomic), I miss Gerrit! ;)

[strohel]

Only thing I might request is showing off new code that this recipient change allows, which wasn't possible before the change.

That's a good point. That will also server as a form of a compile-time unit-test of the API. I've added it in Add example of .recipient() chaining to delay_actor example a772f09e5

With that, I think I'll merge this after my lunch and cut a new release.