Install newrelic_rpm to < 4.0

NewRelic have dropped support for ruby 1.9.x, and so we can't use their latest
app monitoring agent until we upgrade ruby.

ref https://discuss.newrelic.com/t/support-for-ruby-jruby-1-x-is-being-deprecated-in-ruby-agent-4-0-0/44787

TNL-6296: Fixes rake search:rebuild for ficus. (new)

use 16.04 latest docker image for forums

The core issue was that if you don't provide an install step travis
tries to do its default action which generates a bunch of ruby code
that is linked to the ruby on the host system instead of the ruby in
the container.

Add a default response size and a response size limit.

By default, we currently allow a user to retrieve all responses from a
thread, which is terrible from a performance standpoint.

This change adds a default response size of 100, which was chosen to
match typical production pagination parameters, and a response limit of
200.  Given that responses can include children comments, 200 is still a
large number, but far less disastrous than, say, getting 3000 responses,
with all of their children comments.

Both of these values are configurable, and can be overriden via
environment variables.

Add method tracing inside user.

TNL-5115: Remove sort order from cs_comment_service

MA-2139: Mark thread as read for leaner on thread/comment creation

Add an OEP-2 compliant openedx.yaml file

Add the ability to request a thread without its responses.

For cases where we don't need or want the responses on a thread at all,
like getting some metadata about a thread, we should be able to request
it without the responses.  This exists in the ThreadPresenter but didn't
exist as an option that could be passed in the API call itself.

We've added this as an option now -- with_responses -- that defaults to
true, which preserves the existing behavior but does in fact allow
toggling.

This gives us a backwards-compatible upgrade path to allow changes to be
made the LMS so that requests which don't need the responses, or do need
the responses, can be explicit about what they want.

MA-2678: replace use of 'updated_at' for 'read' state

Plucking instead of mapping.

Use pluck instead of map for upvoted/downvoted IDs.

Mapping involves doing the projection locally, whereas plucking will ask
MongoDB to do the projection for us.  This is the difference between
sending the whole document back and sending back the ObjectId itself.

Multipled against hundreds or thousands of documents, mapping can really
add up, so plucking should really help these 99%'ile cases.

Add more tracing + code cleanup.

Revert "Merge pull request #196 from edx/platform/upgrade-to-ruby23"

This reverts commit 22178086, reversing
changes made to 14ed8a85.

Upgrade to Ruby 2.3.0