All textual protocols with legacy of IETF style design use CRLF, and that includes HTTP which heavily drunk from that legacy.
Personally I suspect it's because ARPA was a bit cheap on some things and thus they used typewriters instead of protocol analyzers ;)
Still most if not all implementations use it.
That is alot of bytes when you sum them up.
Also CRLFCRLF is used to demarkate end of headers and chunks.