Fix a crash when trying to get HLS fragments from a disconnected strema.

After the last patch, we would try to serve HLS fragments even if the
backend is down, but we had zeroed out the HTTP header, causing an
assertion failure with HTTP/1.1 clients (and an invalid HTTP response
for HTTP/1.0 clients). Fix by keeping the header and setting a special
“unavailable” flag instead (which allows us to keep sending HLS fragments).