update_attribute regression

I stumbled across this while fixing vestal_versions for a Rails 3 project.

This change breaks the way update_all, dirty tracking and after_save callbacks worked together at least for this particular library (vestal_version) but I'd imagine it breaks a bunch of other implementations, too.

I still wonder what the reasoning behind this change is, but also think that it's a regression and should be reverted:

http://github.com/rails/rails/commit/01629d180468049d17a8be6900e27a...

To demonstrate the issue I've added this test to my vestal_versions fork:

http://github.com/svenfuchs/vestal_versions/blob/master/test/rails3...

The first test passes, the second one fails on AR 3 (where both were passing on AR 2.3.x). Reverting the implementation of update_attribute back to what it was in AR 2.3.x makes the second test pass, too.

If there's an intention of breaking the API in AR 3 and introducing something more consistent for some reason then I'd propose to either

1. make update_attribute behave exactly the same way as update_attributes does
   
2. have another method that does not use validations, callbacks, dirty tracking and pick a name that actually communicates this behaviour

or

1. simply use the same api as save(:validate => true) uses