add linear lr warmup and lr decay scheduler

[lessw2020]

this PR adds a linear lr scheduler and includes some automation based on current best practices:
a - takes user lr provided in args as lr_max, and computes final min_lr for the decay schedule based on lr / 10, per chinchilla paper. (i.e. total decay will be one order of magnitude).
b - computes an automated linear warmup schedule of 10% total iters as warmup, with min warmup of 2 steps.
c - computes a linear decay schedule after warmup, declining from lr_max to lr_min over the end of warmup to end of training. (per Aarons latest paper, linear is preferred schedule).
d - I updated learning rate to 8e-4, in order to provide more visible per iter results to the user assuming debugModel.

LR scheduling produces much improved loss curve:

I added two log prints - the warmup schedule as one line, and then a step and current lr at each iter.
Both could be disabled if too much info.

[wanchaol]

Thanks for adding this. I think we should use the pytorch optim apis instead of doing manual lr update for all optim states. either use LambaLR or LinearLR https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate

[lessw2020]

updates - now uses LambdaLR and does a full linear warmup and linear decay.
Note that LambdaLR does not have args, so for state I had to use two global variables in the lr_scheduling.py.
I updated the train.py args to have a warmup pct (default is 10%), and updated the console display to provide step / loss / curr lr.

[wanchaol]

Sounds good, have a few more suggestion, can either do in this PR or have a follow up.