[fx] add more op patches for profiler and error message for unsupported ops.

[super-dainiu]

What's new?

Now our profiler can profile any model from torchvision library. I also added error message and guide for unsupported ops.

>> model = tm.efficientnet_b0()
>> data = torch.rand(1, 3, 224, 224, device='meta')
>> gm = symbolic_trace(model)
>> MetaInfoProp(gm).run(data)

AssertionError: 
Colossal-AI hasn't supported profiling for <class 'torch.nn.modules.activation.SiLU'>, you might manually patch it with the following code.

from colossalai.fx.profiler import meta_profiler_module

@meta_profiler_module.register(YOUR_MODULE)
def profile_YOUR_MODULE(self: torch.nn.Module, input: torch.Tensor) -> Tuple[int, int]:
    flops = ...
    macs = ...
    return flops, macs

Concerns

This profiler is not able to detect temporary memory cost within one node. For example, nn.Conv2d() has extra memory reserved for the allocator, and the amount varies according to different types of devices (https://dl.acm.org/doi/10.1145/3368089.3417050). Also, I did not consider the extra memory cost when computing nn.MultiheadAttention. Some problems can be solved by refractoring, but we have to withstand the inaccuracies.

[Cypher30]

Great Work!