[fx] add vanilla activation checkpoint search with test on resnet and densenet

[super-dainiu]

What's new?

Now you can use the greedy algorithm from Training Deep Nets with Sublinear Memory Cost (Chen et al. 2016) with two lines of code.

B = 5 * 1024 * 1024 * 1024  # An approximate memory budget of 5GB
model = resnet18()
input_sample = torch.rand(4, 3, 224, 224)
gm = symbolic_trace(model)
MetaInfoProp(gm).run(input_sample)
gm = chen_greedy(gm, B)     # gm will run with activation checkpoint