[Op] Fuse bias+dropout in FusedMLP

[comaniac]

Description

1. Fuse bias and dropout in op.FusedMLP.
2. Add unit tests for op.FusedMLP.
3. Correct random states in GPT schedules when sequence parallelism is enabled.

Checklist

• PR's title starts with a category (e.g. [Bugfix], [Model], [Tutorial], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented

cc @chhzh123 @szhengac

[comaniac]

Based on the comments, I removed the fused bias modules. Instead, I created LinearWithAct and LinearWithDropout. Both modules derive nn.Linear. In addition, I enable memory_efficient_fusion for bias+dropout if available, and add a flag to switch to torchscript in unit tests.

[comaniac]

Thanks @chhzh123 @szhengac