Reduce time cost of grad norm clipping

[danbraunai-apollo]

Grad norm clipping slows down runs by 20-25% for local runs and 5-15% slower for e2e runs. We have not thoroughly tested its value in these experiments.

We can either:

1. Remove it completely. This should be tested by running a small sweep and comparing runs to those reported in the paper (i.e. any run in https://wandb.ai/sparsify/gpt2/.
2. Only use it at the start of training when grad norms are high.