Consider eliminating known-slow requests from request_time stats/histogram

Ran into this with a client recently. Here's what we do right now:

https://github.com/apache/couchdb/blob/master/src/chttpd/src/chttpd.erl#L353-L354

But it might be nice to ignore specific types of requests that are known to run long, that will skew the histogram/averages, such as:

* /_replicate
* any DB and global changes feeds with continuous or longpoll options

We could also relegate these to a separate "long request" histogram stats bucket.

/cc @janl

Start with the request handling and timing code at src/chttpd/src/chttpd.erl#L353-L354, then trace how request_time statistics and histograms are recorded. The scope needs a decision between excluding replication and continuous or longpoll feed requests or placing them in a separate long-request bucket; done means the chosen behavior is defined and reflected in those stats.

Written by the indexing model from the issue text.