Add fp8 scaling, fp8 scaled convolutions, and gradient checkpointing for Flux AutoEncoder

[rockerBOO]

I am working with some processes that use VAE as part of the gradient flow, so need forwards/backwards support for decoding specifically. But this causes the gradient to build up significantly inside the AutoEncoder. This PR works to add fp8 scaling to the AutoEncoder and gradient checkpointing to help reduce how much VRAM this uses.

• Add fp8 scaling to Flux AutoEncoder
• Add gradient checkpointing to Flux AutoEncoder for encode/decode