Don't call self.inner.read() with empty buffer

[mdsteele]

Alas, I have just learned the hard way that some Read implementations (in this case, flate2::read::GzEncoder) don't behave nicely when read() is called with an empty buffer (you'd think they'd just return Ok(0), but GzEncoder gets confused). However, the Read implementation for newc::Reader calls self.inner.read() with a buffer limited in length to the size of the remaining data, which could be zero even if the original buffer was nonempty. (And so I had a bug where copying data from an entry in a gzipped cpio archive would fail.)

This PR adds the simple workaround of not bothering to call self.inner.read() if limit is zero. I think this is probably slightly better anyway, even for better-behaved inner readers, since self.inner.read() could potentially involve a syscall or some other relatively expensive operation, and there's no reason to do that if we already know we won't be returning any data.