Let IO#sync only affect writes, introduce IO#read_buffering? and use it in urandom

[asterite]

Fixes #6303

Being able to disable read buffering for IO::Buffered is good. However, having sync affect both reads and writes makes it easier to enable just buffering for one direction: for sockets it's better to not have write buffering by default, so whatever you write goes directly to the socket (you can always set sync = false and manually flush, which is what we do in a few places), but reads should always be buffered for performance reasons.

With this PR we could release 0.25.2 as a bugfix release for this performance regression.

Then for 0.26.0 we could rename sync to write_buffering, or just leave it as sync. Maybe write_buffering is more explicit, even though it's a breaking change. But then, disabling read buffering is a very uncommon use case (urandom... what else?), so maybe having sync and read_buffering as separate, asymmetric properties isn't that bad.

(I know I could have used getter or setter for this, but given that sync doesn't use them I decided to follow that style. If someone wants to do a follow up PR to "improve" that (it's not really an improvement, because macros are slower to compile and it's just one/two line differences), then please go ahead).

[asterite]

By the way, to review this you can see 4e0dc9e. It's more or less the same code, except that when using sync for reads it now uses !read_buffering?

[Sija]

As this PR fixes serious perf regression IMO it should go into 0.25.2.