gradients through where when one of the branches is nan

What is tinygrad's plan to deal with gradients of functions defined with a Tensor.where operation in which the branch not taken is nan but the gradient still comes out to nan due to 0*nan resulting in nan. The issue is present in jax as well and is described in detail here - https://github.com/jax-ml/jax/issues/1052#issuecomment-514083352.

It can be reproduced in tinygrad with something like

from tinygrad import Tensor

x = Tensor([0], dtype="float32", requires_grad=True)
y = Tensor.where(x == 0, 0, x * x.sqrt())
y.sum().backward()
print(x.grad.tolist())

As they mention, the right solution which "is pretty heavyweight", "is something like tracking a symbolic zero mask potentially through the entire backward pass of differentiation".

Start by running the provided Tensor.where reproduction and compare its gradient behavior with the linked JAX discussion. Trace the backward pass for Tensor.where and determine the intended handling of a NaN in the unselected branch; done should include a documented plan and a verified result for this example.