1. 1. 1. 🐛 Describe the bug

I'm observing a memory leak when using a custom torch.autograd.Function with an in-place operation marked by ctx.mark_dirty(...). The following minimal repro highlights the issue:

```python
import torch

class ReluOps(torch.autograd.Function):
@staticmethod
def forward(ctx, x):
x.relu_()
ctx.mark_dirty
ctx.save_for_backward
return x

@staticmethod
def backward(ctx, grad):
x = ctx.saved_tensors
return (grad * x[0])

def run():
for i in range(100):
x = torch.rand((100, 100, 1000), requires_grad=True)
z = x + 1.0
z_view = z[0]
ReluOps.apply(z_view)

if _name_ == "_main_":
run()

```

Memory usage increases over iterations, and is not released between invocations.

The cycle looks like this:
CopySlices -> PyNode -> THPFunction -> SavedVariable -> AsStridedBackward -> CopySlices

1. 1. 1. Versions
         torch master
         run above code

cc @ezyang @gchanan @zou3519 @kadeng @msaroufim @albanD @gqchen @nikitaved @soulitzer @Varal7 @xmfan @chauhang @penguinwu