Revert "Make scopes available for abstract classes"

[paracycle]

Reverts #1250

I've been thinking about the problem that this change has been causing with AR relations (as pointed out by #1413 that was trying to fix this forward), and I've come to the conclusion that it is better for Tapioca to NOT support this case at all, for a few reasons.

I will explain my reasons based on the following code example (adapted from the one in the original PR #1250)

ActiveRecord::Schema.define do
  create_table :posts, force: true do |t|
    t.datetime :deleted_at, nilable: true
  end

  create_table :comments, force: true do |t|
    t.integer :post_id
  end
end

class ApplicationRecord < ActiveRecord::Base
  self.abstract_class = true

  scope :deleted, -> { where.not(deleted_at: nil) }

  def self.count_deleted
    deleted.count
  end
end

class Post < ApplicationRecord
  has_many :comments
end

class SubPost < Post
end

class Comment < ApplicationRecord
  belongs_to :post
end

1. Given the current level of RBI signature generation there is no good way to type the return value of the deleted method. It is supposed to the an ActiveRecord::Relation subtype, but the actual type would depend on what it was called on. For example:

   Post.deleted.class    #=> Post::ActiveRecord_Relation
   SubPost.deleted.class #=> SubPost::ActiveRecord_Relation
   Comment.deleted.class #=> Comment::ActiveRecord_Relation

   and we are not able to currently type this appropriately.
2. More importantly than the previous point, there is not generic way to type the deleted call being made inside the count_deleted method, since depending on what model type count_deleted is being called on, it could be returning any of the types listed in the previous point. As a result, the deleted call being made inside that method is not type safe anyway, and I would really prefer if that was explicit using a T.unsafe explicitly.
3. Finally, and probably, most importantly, I think the pattern that is being suggested is unsafe to begin with. You can easily see that by trying to make a Comment.count_deleted call and seeing that it raises an error with the message no such column: comments.deleted_at. Since the abstract base class cannot enforce the definition of a set of columns on child models, it should not try to act like it can ensure that it can make these kinds of queries either. If applications still want to use this pattern, it is better for them to explicitly use T.unsafe to communicate that the method could be unsafe in this respect as well.

So, I think it is best to revert the original PR and disallow abstract base classes from DSL generation for scopes.