[FEATURE]: Meta information patch for `torch.matmul`

[Cypher30]

Describe the feature

To provide strategies with our auto-parallel system, the meta information (including compute cost and memory cost during the training process) is the key factor in generating a good strategy. We have done a lot of work previously, including the MetaTensor #1515 mechanism. With the MetaTensor mechanism, we are able to access the lower-level operations during the training process. Then we construct profiler #1587 on top of this mechanism. However, we found that the profiler itself might be time-consuming when inspecting all the operations in the model. Therefore, we decided to make use of MetaTensor to retrieve the Aten-level graph and turn all the meta information calculation into the static formula, so that we could avoid calling torch.autograd.backward every time we want to profile a PyTorch operation. We have done a lot so far and manage to combine the SPMD solver and auto activation checkpoint solver on ResNet #2258.

In order to combine the SPMD solver and auto activation checkpoint solver on transformer-based networks, we need to patch the torch.matmul operation.