That only applies to the headers, not the body of the response. If the Transfer-Encoding is "identity" or omitted (in which case I believe "identity" is implied), the line endings can be anything; the body is just a stream of octets. If the line ending had to be CRLF in the body, what would that do to binary downloads?
Even if the RFC can be interpreted to say that the message body MUST have CRLF line endings (which I doubt), in real life, documents served by HTTP servers have all sorts of line endings (I've even seen mixed line endings in HTML files delivered via HTTP). If you think otherwise, look at the bytes delivered from http://www.microsoft.com/ (CRLF), http://www.linux.org/ (LF), http://www.linux.com/ (mostly LF with a few errant CRLFs), and (drat... couldn't find any pure-CR URLs).