Update on "[ET] enabling half dtype output for dequantization and making logic consistent"

Enabling half dtype output and making dequantization logic consistent between per_tensor and per_token as it is currently prone to integer overflows on one over the other

Differential Revision: [D76289181](https://our.internmc.facebook.com/intern/diff/D76289181/)

[ghstack-poisoned]

Author: morelos