Stop aggressive inlining.

It's not clear what exactly is happening here, but the Read implementation
for text decoding appears a bit sensitive. Small pertubations in the code
appear to have a nearly 100% impact on the overall speed of ripgrep when
searching UTF-16 files.

I haven't had the time to examine the generated code in detail, but
`perf stat` seems to think that the instruction cache is performing a lot
worse when the code slows down. This might mean that excessive inlining
causes a different code structure that leads to less-than-optimal icache
usage, but it's at best a guess.

Explicitly disabling the inline for the cold path seems to help the
optimizer figure out the right thing.

src/decoder.rs

@@ -251,6 +251,7 @@ impl<R: io::Read, B: AsMut<[u8]>> DecodeReader<R, B> {
         Ok(nwrite)
     }
 
+    #[inline(never)] // impacts perf...
     fn detect(&mut self) -> io::Result<()> {
         let bom = try!(self.rdr.peek_bom());
         self.decoder = bom.decoder();