[assoc_scan] Memory-efficient backward via associative_scan (#192761)

This PR replaces the matrix-based gradient computation of associative_scan with a backward associative_scan. It enhances readability, maintainability and memory footprint. However, runtime needs to be checked.

It replaces #162906 and has been consolidated into a ghstack with the vmap implementation.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/192761
Approved by: https://github.com/ydwu4