- Increase the maximum response header buffer size up to ~32 kB.
  This also moves it into a heap-allocated buffer instead of the
   stack, as 32 kB is really too much for the stack.
- Treat LF as CR LF when parsing headers.
  This is necessary when parsing the load-balancer response in
   front of <http://lazaradio.hu:8100/bermuda.opus>.
  The response returned by that server is invalid in lots of ways,
   but with these two changes we can read it.
- In addition, we now peek ahead at a large chunk of data when
   reading the response instead of reading 2 to 4 bytes at a time.
  This allows a typical response to be read with two syscalls
   (one peek, one read) instead of several hundred.
- Stop trying to read more data when the connection is closed.

The big one was that if the connect() call failed, it would loop
 forever (thanks to some code re-factoring, the loop was no longer
 advancing to the next address as originally designed).

On the first request made by a connection, the value it was testing
 might be uninitialized, so we can't guarantee it'll pass.

* Now supports HTTP/1.1 persistent connections with pipelining.
  This speeds up chain enumeration on large files by almost a
   factor of 2 over http, and by roughly a factor of 4 over https.
  The difference between http and https is now much smaller.
* Add timeouts to all the socket I/O.
* Estimate the number of available bytes to read and use it when
   making connection re-use decisions.
* Add support for https with proxies using HTTP/1.1 CONNECT
   tunnels.
* Fix TLS session re-use (it requires clean shutdown).
* Various other code re-organization and minor improvements.

* Tags were being freed even if not initialized if
   op_fetch_headers() failed on the first link.
  The logic for handling completely empty links would leak, also.
* Ignore fragment identifiers in http URLs instead of rejecting
   them.
* Get the current link in seeking_example's verify_seek when
   op_read_native() fails to return it for us.

Most of these were dead code left around from rewriting things and
 things llvm doesn't have enough global information to prove by
 itself.
The one real error was the missing NULL check in
 opus_tags_parse_impl().

The previous comparison would have succeeded on things like
 "localhost123" instead of just "localhost".

* s/op_read_stereo_float/op_read_float_stereo/ for the fixed-point
   API.
* Fix compiler warnings exposed when optimizations are enabled.
* Fix opusfile_example to work with --enable-fixed-point
   --disable-float
* Fix seeking_example to not re-define OP_FIXED_POINT if it's
   already been defined.

The most important for the https backend:
- Implement RFC 6066 Server Name Indication.
- Implement SSL session resumption.
  This gives less of a speed improvement than you might think.

Teach me to change something right before committing.

This gives us a better idea when to re-use a connection.

Actually close the old socket.
Don't initialize SSL more than once.

This is pretty preliminary for now.
Seeking performance isn't great, and there's no caching.
I'm sure there's also lots of bugs.