Loss format from .3f to .3g in the training loop

[fjhheras]

🚀 Feature

I propose to change the default format of loss during training (in the tqdm bar) from .3f to .3g

Motivation

When using pytorch-lightning with losses that are quite close to zero, the tqdm information during training becomes non informative, because the loss is always 0.000. For example:

Epoch 41:  76%|██████                                   | 37/49 [00:00<00:00, 65.74it/s, loss=0.000, v_num=92]

The proposed change takes it to

Epoch 41:  76%|██████                                   | 37/49 [00:00<00:00, 65.74it/s, loss=2.2e-08, v_num=92]

This change from .3f to .3g has also an advantage when loss is a large number

$print('loss: %.3f' % 80808423243)
loss: 80808423243.000
$print('loss: %.3g' % 80808423243)
loss: 8.08e+10

In other situations, the output of .3f and .3g does not change

$print('loss: %.3f' % .884)
loss: 0.884
$print('loss: %.3g' % .884)
loss: 0.884