Lower cummax op

[zyy-martin]

Implement cummax op. Fixes #8230

Cummax returns a tuple of (values, indices). This PR uses the ReduceWindowWithGeneralPadding XLA op (padding + window reducer for cumulative values) to implement it similar to cumsum/cumprod. The only difference is that the indices will be passed to the ReduceWindowWithGeneralPadding op along with the values, and a new reducer is implemented to look at both the value and index of input and return a tuple of (max_value, max_index).

[zyy-martin]

I exclude the cummax from the python op test because I believe the scalar value generated from the sample input would fail, as will existing ops like cumprod and cumsum. The following script will fail for example:

import torch
import torch_xla
import torch_xla.core.xla_model as xm

a = torch.tensor(7.358110427856445, device=xm.xla_device())
print(f'[xla]: {torch.cumsum(a, 0)}')

[zyy-martin]

Thanks for the review! Can someone with write access merge the PR? Thanks!