Updating LAMB optimizer to v3

[Shashi456]

While LAMB v1 had the debiasing step as follows : LAMB v3, no longer comprises of the same:

Also even pre-PR, I observed that

let num = debias1 * state[StateKeys.avgGrad]!

while it should've been :

let num =  state[StateKeys.avgGrad]! / debias1 

since the pythonic version, would look something like:

 step = (exp_avg/debias1) / ((exp_avg_sqr/debias2).sqrt()+eps)

[sgugger]

Thanks!