Add Send marker trait to into_reader.

[jsha]

This allows the resulting Read to be shared among threads.

[algesten]

Makes a lot of sense.

I'm unsure about Sync here. That means we guarantee the reader isn't only possible to pass around to other threads, but it's also consistently observable simultaneously in all threads.

In practice our impl maybe always wrap the actual thing in a mutex, so we do guarantee that. I'm thinking maybe we some future impl of our reader has some sweet optimization we can't uphold with Sync which means a breaking change. But maybe that's too much thinking?

Since we don't impl Clone and read() call is using &mut self, in practice it maybe doesn't make much difference?

@rawler maybe has some input? :)

[jsha]

Pushed a change to narrow this to just Send, which is what I need to unblock some work on git2-rs.

In practice it looks like our Read gets Sync because TcpStream, StreamOwned, and Cursor all get it. My thinking was that with dyn, once we've "cast away" the marker traits, it's impossible for user code ever to get them back (as opposed to generics, where the marker traits automatically propagate through). In that context it felt a shame to throw away properties that might be valuable to the user. But I do see your point: the impl Read return hides some of our implementation details (like ChunkDecoder, DeadlineStream, PoolReturnRead) from the user, which gives us flexibility; and one of the things we might want to keep hiding is the Sync'ness so we have flexibility there.

[algesten]

Perfect. Let's merge!