[Refactor] Schedule primitives

[comaniac]

Description

This PR refactors the schedule primitives to make them more systematic and extensible:

• Now all primitives derived the base class Primitive that defines the required APIs (name, apply, and optional init_metadata).
• The apply function takes the schedule as the first argument, followed by any primitive specific arguments.
• The init_metadata allows each primitive to register their metadata, so the original schedule metadata only defines tie_weights metadata now. Meanwhile, we currently also put some metadata, such as the original shape before sharding, directly to nn.Parameter. I'll send a follow-up PR to better organize them.
• All primitives are registered and maintained in a table, and they will be dynamically set to a schedule when it is constructed. In other words, users can register their primitives outside of the Slapo code base.

In addition this PR also

• Move build related functions to build.py.
• Rename get_output_type_after_sharding to postproc_sharding, because this function not only infers the output type but also changes the attributes (e.g., in_feature) after sharding.

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

[chhzh123]

Otherwise LGTM

[comaniac]

Thanks @chhzh123