[fx] Add activation checkpoint solver rotor

[Cypher30]

What's New?

In this PR, we introduce rotor algorithm to our activation checkpoint solver lists. By the great effort in #1480 , we are able to estimate the metrics we need to use rotor algorithm. Down below is a quick example to try this new feature

import torch
import torch.nn as nn
import torch.nn.functional as F
from colossalai.fx import ColoTracer, ColoGraphModule
from colossalai.fx.passes.algorithms import linearize, solver_rotor
from torch.fx import GraphModule
from torchvision.models import resnet18
from colossalai.fx.codegen import ActivationCheckpointCodeGen

model = torch_resnet18()
input = torch.rand(32, 3, 224, 224, device="meta")
tracer = ColoTracer()
graph = tracer.trace(torch_model, meta_args={"x": input})
graph.set_codegen(ActivationCheckpointCodeGen())
gm = ColoGraphModule(model, graph)
gm = solver_rotor(gm, input, mem_limit=600 * 1024 * 1024)
gm.to_folder("./fx")

I also add the unit test for it in test_ckpt_torchvion.py, currently on our CI we will skip this test, as we haven't implemented our ColoGraphModule in torch11, I will attach the test results on torch12 later.

[Cypher30]

Above the test result in `test_fx` with torch12