stop assuming bwd = 2x fwd

[jfc4050]

this is close enough assumption a lot of the time, but it starts falling apart for cases with high model parallel comm. for example expert MLPs backward has wgrad and dgrad, but still only has 2x alltoall

[jfc4050]

first pass here https://github.com/jfc4050/dlcalc/tree/no-bwd-approx