[context] support lazy init of module

[FrankLeeeee]

The current ColoTensor does has not supported lazy initialization yet. This PR proposes a method for lazy initialization and it's currently experimental and subject to future changes.

In this PR, a context is implemented for lazy initialization, this context will do the following things.

1. hijack the module initialization so that all modules are created as meta tensors. This can have two benefits 1) super fast module initialization 2) make the module traceable for future integration with torch.fx
2. hijack the weight initialization calls and put them into cache
3. materialize the meta tensors to real-value tensors and apply weight initialization when reset_parameters is called.