return an opaque type for time zone name abbreviations in `TimeZone::to_offset`

[BurntSushi]

For similar reasons in #221, I believe it is more flexible to use an opaque type for the time zone abbreviation returned by TimeZone::to_offset instead of a &str. In particular, the &str requires that the abbreviation be borrowed from the TimeZone. In core-only environments, this in turn implies that a fixed offset time zone (the only type of time zone available in such environments at present) has to be converting to a string and stored inline. But if we returned an opaque type, then I believe we could do this string conversion lazily instead of making every TimeZone value pay for it.

This opaque type would still need to be parameterized over a lifetime borrowed from the TimeZone so that we can borrow data from the TimeZone. But in core-only environments, I think this would end up being, effectively, a phantom lifetime.

[BurntSushi]

I ended up splitting to_offset into to_offset and to_offset_info, with the latter expected to be called extremely rarely while the former is called extremely frequently. The former now just returns the offset while the latter includes the DST status and the time zone abbreviation. This freed me up to do a little more work to produce the time zone abbreviation in the fixed offset case. And this turn let me remove it from the TimeZone itself. This overall reduced the size of TimeZone from 20 bytes to 8 bytes on x86-64 in core-only environments. And consequently reduced the size of Zoned as well.

Also, the lifetime of TimeZoneTransition::abbreviation has been made more restrictive to make future semver compatible API evolution a bit more flexible.