It used to only read the data it could get without blocking. Andreas says this
was intentional, but it's counterintuitive and no code that uses read_all()
actually expects it to return only a part of the data. So change it to always
read data until an EOF (or an error) is received.
We were returning a zero-length ByteBuffer in some cases. We should be
consistent about this and always return a null ByteBuffer if nothing
was read at all.
There's some confusion between the write syscall and CIODevice::write()
here. The internal write() returns a boolean, and has already whined
in case the syscall failed, so we don't need to do that again.
Add a trivial CSafeSyscall template that calls a callback until it stops
returning EINTR, and use it everywhere we use select() now.
Thanks to Andreas for the suggestion of using a template parameter for
the syscall function to invoke.
