Can you add tests confirming that this is okay to do, and things behave as expected?

This branch currently is behind tinygrad/master. The line count difference bot is disabled.

Added some simple tests. Let me know if you have better ideas as I can't think of any now.

Also, I don't fully understand what .assign() does, how it's different from .replace(), and why they have to be separate.

I mean test that gradient behaves as expected, not that the flag is right!

replace/assign seems to work differently from torch.Tensor.copy_ (see my test). All the optimiser tests are failing when requires_grad is set, as they use assign. Should replace/assign be separate?

closing as stale