Downsampling uses ordered grids. Hardware-supported AA can use rotated and otherwise sparse grids, and gives you bigger quality gains per sample.
Resolving deferred rendering buffers on multiple samples per pixel is completely trivial, and not in any way slower than doing it on a brute-force oversampled buffer. Actually, the same rule applies, in that you have fewer total samples in your buffer, and your resolve pass will be faster accordingly.
There may be some backwater engines that just can't do this properly, and of course it's a valid strategy to use brute force to compensate for engineering incompetence. But you should always check first to see if it works the proper way. You throw way less computing resources out the window this way.