feat(rt): add Timer trait

[seanmonstar]

This adds a hyper::rt::Timer trait, and it is used in connection
builders to configure a custom timer source for timeouts.

Closes #2846
Closes #2847
Closes #2848

Continues #2904

[seanmonstar]

Continuing the conversation, where @dhobsd said:

Thanks for remembering us and reaching out! (Sorry for delay!)

Fuchsia's async executor uses deadlines for its time, so that's at least good. Two points to bring up:

• reset in tokio says "This function can be called both before and after the future has completed." In fuchsia_async's Timer, this isn't true -- you can only reset a timer that's already fired. I'm also not sure what it means to reset a timer that hasn't fired because it's hard to guarantee you won't observe the signal. Does hyper reset timers that haven't fired?
• is_elapsed says that it returns true when the time has elapsed, but is this the same as saying the timer has already fired?

I guess a third question: will hyper rely on these divergent behaviors? If not, perhaps we can make the documentation stricter in hyper's trait?

[seanmonstar]

So I looked around how we use these methods, and here's what I've found:

• reset is used for the HTTP/1 header read timeout, so that successive reads simply move the timer further out, instead of allocating a new one.
• Doesn't look like we use the other methods, so I've been happy to toss them.

The reset method seems to just be to allow for optimizations, but we could simply provide a default implementation of the method somehow... Basically, if it didn't exist, we'd just call timer.sleep(next_dur) and replace the old one. Hm, that might mean the method needs to be on Timer instead, since we can't make a default implementation that creates a new sleep without the Timer. Maybe then a timer.reset(sleep, next_dur).

We even could just leave it off entirely to start with, since adding a method with a default implementation is fine. It would just mean potentially unnecessary allocations when using Tokio.

[seanmonstar]

Ok, I've done that: moved the reset method to Timer, and it has a default implementation that just allocates a new sleep.