Fix total_count when grouping model has default scope

[eitoball]

This PR should fix #total_count when grouping model has default_scope.

Currently, if you have model like:

class User < ApplicationRecord
  default_scope -> { where(deleted_at: nil) }
end

and have statement like User.group(:email).having("COUNT(name) >= 3").page(1).total_count, you will get error like "Mysql2::Error: Unknown column 'users.deleted_at' in 'where clause'".

#total_count in 1.2.0 uses subquery (here). This remove table reference from FROM clause, and WHERE clause created by default_scope loses table to reference.

This PR removes default_scope by calling unscoped to solve this issue.

[ChrisBAshton]

This PR also fixes our issue: alphagov/whitehall#5300

In our case, we're not using a default scope but a custom one:

def unfiltered_scope
    StatisticsAnnouncement.includes(:current_release_date, statistics_announcement_topics: :topic, publication: :translations, organisations: :translations)
                          .joins("INNER JOIN statistics_announcement_dates
                            ON (statistics_announcement_dates.statistics_announcement_id = statistics_announcements.id)")
                          .joins("LEFT OUTER JOIN statistics_announcement_dates sd2
                            ON (sd2.statistics_announcement_id = statistics_announcements.id
                            AND statistics_announcement_dates.created_at > sd2.created_at)")
                          .group("statistics_announcement_dates.statistics_announcement_id")
                          .page(options[:page])
end

[bendilley]

This is a good catch and should definitely be merged IMO. I've opened #1015 which covers a variation of the same problem (when STI subtypes get involved) and which needed an additional unscope(where: :type). I've incorporated this fix so both PRs can be merged with a more obvious conflict resolution.

[yuki24]

Hmm, I think calling unscope would be very surprising. I think we should just revert #979. Also the tests in this PR and #1015 are very useful to catch failures early, we might as well want to do that once we revert that.

[bendilley]

This is unscoping c.model in order to say from(subquery_with_all_the_scoping).count so within the remit of acquiring the correct count for a grouped query, I'd say the unscoping is not only ok, but the only way to be sure of deriving the correct total.

[yuki24]

I have reverted #979 and picked up the test from this PR. Thanks for reporting this issue!