1. No dict proxies
2. No dataclass proxies

The main motivation is that fx passes are really difficult to write when there are dict or dataclass operations in the graph. E.g. the dict lookup below

For dicts, this means we would rewrite the triton_kernel_wrapper* HOPs to not use dicts, or shepherd them to use flat_apply HOPs.

For dataclass, we would need to fix invoke_quant somehow.

Thoughts @oulgen, @aakhundov, @davidberard98, @eellison?

cc @chauhang @penguinwu @ydwu4 @bdhirsh