float8 investigate cuda graphs + dynamic scaling leading to memory fragmentations

[vkuzo]

TODO add more context / a repro, for now just capturing the request from chat

copied from meta-pytorch/float8_experimental#249