Fix numericality validator to still use value before type cast except Active Record

[kamipo]

The purpose of fe9547b is to work type casting to value from database.

But that was caused not to use the value before type cast even except
Active Record.

There we never guarantees that the value before type cast was going to
the used in this validation, but we should not change the behavior
unless there is some particular reason.

To restore original behavior, still use the value before type cast if
came_from_user? is undefined (i.e. except Active Record).

Fixes #33651.

[edzhelyov]

While this might close #33651 it won't affect the regression that I have when using ActiveAttr.

This is basically what I have:

class User
  include ActiveAttr::Model

  attribute :account_crn, type: Integer
end

describe User do
  it { should validate_numericality_of(:account_crn) }

  it 'validates account CRN numericality' do
    user.account_crn = 1.1

    user.validate

    user.errors.should be_added :account_crn, :not_an_integer
  end
end

The shoulda-matchers test is broken and my custom one. I will work out to provide a runnable example tomorrow.

[edzhelyov]

I've filled an issue #33686 with the problems that I have after fe9547b. Which this PR won't fix.

[cgriego]

For maximum backwards compatibility, if #{attr_name}_came_from_user? isn't defined then ActiveModel should use #{attr_name}_before_type_cast like it did before instead of read_attribute_before_type_cast.

The read_attribute_before_type_cast method is an ActiveRecord interface, but the ActiveModel convention (using ActiveModel::AttributeMethods) is that #{attr_name}_before_type_cast is the public interface backed by an attribute_before_type_cast method. ActiveRecord makes attribute_before_type_cast private and exposes its own read_attribute_before_type_cast publicly in addition to #{attr_name}_before_type_cast.

[kamipo]

You are right. I've restored to use #{attr_name}_before_type_cast in this method.