enhancement bert training op clustering

To utilize BladeDISC accelerating bert training, we need to cluster more op into disc, as the experiment, we found 244 `FusionGroups` in the first iteration with LTC, to reduce the clusters, we should:

- enhancement op coverage, #441 lists all ops that required, I just list some priority ones:
- native_layer_norm
- _softmax
- _log_softmax
- _log_softmax_backward_data
- _nll_loss_backward
- _nll_loss_forward
- tanh_backward
- native_dropout_backward
- gelu_backward
- enhancement shape propagation
- native_layer_norm
- native_dropout
- native_dropout_backward

[here](https://gist.github.com/Yancey1989/2ffbdea3c48baa0243e56d436eecfc98) list the 224 FusionGroups