Arvados

• Instead of adding a Do field to keepclient.KeepClient as described, I changed the Client field's type from *http.Client to a HTTPClient interface. Otherwise, using the Do field would make the Client field meaningless, except that the "set appropriate timeout according to proxy/disk services" code would edit it uselessly.
• In order to make the "loop detected" error visible to the outer keepproxy on writes, I had to adjust the keepclient error handling. Now, instead of just returning a generic "insufficient replicas" error object, it returns the errors returned by the backend servers, like we do for reads. I also updated the error identifiers to match Go conventions ("FooError" is a type, and "ErrFoo" is an object).

• Requests initiated by keepproxy have a "Via: HTTP/1.1 keepproxy" header where HTTP/1.1 is the protocol of the original request.
• Responses returned by keepproxy have a "Via: HTTP/1.1 keepproxy" header. It's possible the actual upstream protocol (between keepproxy and keepstore) is HTTP/2 but this distinction isn't exposed by the keepclient interface, and seems unimportant anyway.
• When a loop is detected, the client receives a 500 response and the server log says:

  2017/04/27 15:24:49 proxy loop detected (request has Via: "HTTP/1.1 keepproxy"): perhaps keepproxy is misidentified by gateway config as an external client, or its keep_services record does not have service_type=proxy?
  

11537-keepproxy-loop @ baa1c3db1f635828dda3cebdc47d98b33bb6518d

11537-keepproxy-loop LGTM