[Example] Use .fuse() primitive when possible

[chhzh123]

Description

This PR fixes some issues for the example models:

• Use .fuse() to fuse the bias+GeLU in the MLP module. Since TorchScript module cannot be hooked and cannot properly work with DeepSpeed ZeRO-3, this feature is unset by default.
• Use subgraph matching to only optimize a small part of the Albert model instead of replacing a whole layer
• Fix Albert attention mask shape
• Remove useless functions in the schedule
• Fix an argument issue in .replace() and support bias+layernorm fusion

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

[chhzh123]

I didn't change the GPT example to avoid conflicts. You can merge #41 first, and I will update other models correspondingly. @comaniac

[comaniac]

Otherwise LGTM

[comaniac]

LGTM

[chhzh123]

Please do not merge first. I'll also remove the epoi dependency.

[chhzh123]

Or do you think it is better to create a separate PR? @comaniac

[comaniac]

I'm fine with both, so just for your convenience.

[chhzh123]

I've changed all the epoi.attention to slapo.op.attention (except for the T5 one). This PR is ready to review and merge @comaniac.

[comaniac]

Thanks @chhzh123