Decouple quantization and activation dtypes for non-xnnpack

[jackzhxng]

🚀 The feature, motivation and pitch

After #8488 gets in, which allows us to decouple the dtype used at weight quantization time and the dtype used during op computation for XNNPack 8da4w only, enable this for other backends and quantization settings as well. This should be possible with the new quantize_ API.

cc @mergennachin @iseeyuan @lucylq @helunwencser @tarun292 @kimishpatel @andrewor14

Alternatives

No response

Additional context

No response

RFC (Optional)

No response