Features from LLMR - config params

[madragonse]

features present previously in LLMR

[madragonse]

Parametrization of lr scheduler - warmup lr start fraction

[madragonse]

Save model at the end of the training/ft - config feature switch/config

[madragonse]

Disable periodic checkpoint save

[madragonse]

warmup_steps should be in total steps % or both options available

[crewtool]

LR in log scale - mitigates issue with long and misleading representation and sets common discrete lr steps values in more approachable form

You can change to log scale in neptune

[madragonse]

was thinking about exp(-xlog2) representation of lr, already added #66