[ENHANCEMENT] Enable non-gelu activations for BERT LM Head

[skothenhill-nv]

Is your feature request related to a problem? Please describe.
Bert LM-Head hardcode's gelu as the activation, this has caused problems in downstream projects (e.g. NeMo).

Describe the solution you'd like
Rather than hardcoding nn.gelu, we accept the field set to TransformerConfig.activation_func as done for other models.

Describe alternatives you've considered
did not consider alternatives

Proposed implementation
PR.

Additional context
Spoke with @shanmugamr1992 and used his suggested solution.

[github-actions]

Marking as stale. No activity in 60 days.