Estimate the Adahessian trace on both complex components

[shaneraphel]

Summary

The Adahessian half of #458. torch.randint_like rejects a complex parameter, so get_trace raised RuntimeError: check_random_bounds handles only integral, floating-point and boolean types and the step never ran. AdaBound's complex failure in the same issue is a different call and is handled separately.

The Hutchinson vector is now a Rademacher draw on view_as_real of the parameter, packed back with view_as_complex so it still matches the complex gradient that autograd produced. The absolute value that removes the ±1 sign is taken per component. A complex conv kernel still averages its spatial dimensions, as the real 4D path does; the component axis is not one of those dimensions. Moments are stored on the two components and the write-back goes through view_as_real, so neither part is dropped.

A real parameter does not take this branch. Two steps on [0.5, -0.2, 0.3] still finish at [0.43060994148254395, -0.17224396765232086, 0.2583659589290619].

Complex rank 3 and 5 are still rejected. The real implementation has no branch for those ranks either, and a complex conv1d or conv3d would need the same spatial reduction the real kernels are still missing.

Test plan

• pytest tests/test_adahessian_complex.py
• Real path matches the previous two-step result.
• One step on [0.5+0.1j, -0.2+0.3j] matches Adahessian on view_as_real of the same values, bitwise, including the random signs.
• A 2×2 complex matrix takes a step and stays complex64.

This change was produced with assistance from an AI coding tool. I reproduced the randint error, checked the complex step against the real algorithm on both components, and ran the tests locally.