Bias HTTP detection towards availability

[olix0r]

1b97e57 changed HTTP detection to bias towards correctness. Instead of
performing a single read, the HTTP detection module reads data until a
newline is encountered so that a dispositive determination can be made.
Practically, this causes many non-HTTP protocols to fail with a
connection timeout because a newline is never transmitted.

This change reverts the HTTP detection logic to perform a single read
and to use that data to make a determination. A minimum of 14B is
required to handle a connection as HTTP.

[hawkw]

This change reverts the HTTP detection logic to perform a single read
and to use that data to make a determination. A minimum of 14B is
required to handle a connection as HTTP.

As we discussed offline, it seems like a potential compromise (which would require a somewhat bigger code change) would be to make multiple reads, but fall back to transparent forwarding when detection times out, rather than failing the connection. This would let us avoid false negatives in the case where a single read doesn't return enough data to make a protocol determination, but would still bias for availability. I'm not sure whether or not this is necessary though. It seems pretty unlikely that we'll need multiple reads to detect HTTP.

We may want to do this in a follow-up regardless, but this change seems good to me for now.