Hi,

I'm working on a small webserver; now that we have the current_thread runtime in tokio it seems interesting to write thread-local async code.

This would allow us to use "unsync" versions of Bytes and BytesMut; the idea was raised in #99 before.

The overhead is in the atomic operations; thread-local versions can use a simple ref_count: usize, and also don't need to worry about concurrent calls of shallow_clone (although by making Bytes and BytesMut !Sync that could be avoided too; they'd still be Send).

I have a work-in-progress fork at https://github.com/stbuehler/bytes/tree/unsync if someone wants to try running a few benchmarks (I basically reimplemented it from scratch, without original_capacity. I also added a UnBytesExt variant that defaults to extending the allocation).

In the existing benchmarks the "unsync" version wins on my system mostly the split_off_and_drop case (#118 might help the "sync" version on that), and some small improvements for drain_write_drain, fmt_write and from_long_slice - but loses in defer_* and slice_* (and I have no idea why; in disassemble dumps it looks like the compiler favors the inline branch in the "sync" part but not in the "unsync" part; exported as_ref wrappers looked exactly the same otherwise).

It would also be possible to put a shared reference to a "sync" Inner into the "unsync" Shared, so the "unsync" types could also refer to storage shared between threads on demand.