Run AdaBound on the real and imaginary parts

[shaneraphel]

Summary

Fixes #458 for AdaBound. step_size.clamp_ is not implemented for complex tensors, so a complex weight raised NotImplementedError: clamp is not supported for complex types at the learning-rate bound and never updated.

A complex parameter is updated as two real components. The moments, the bound and the write-back run on torch.view_as_real, which is how Adam updates a complex weight: both the real and the imaginary part are kept, and neither is dropped. Real parameters take the same statements as before.

Adahessian's complex failure in the same issue is a different call (torch.randint_like on a complex tensor) and is not changed here.

Test plan

• pytest tests/test_adabound_complex.py
• Two real steps with weight decay match the previous implementation exactly, including the value 0.09872793406248093 on the first coordinate.
• A complex step, and the AMSBound variant, match AdaBound run on view_as_real of the same parameter and gradient, bitwise.

This change was produced with assistance from an AI coding tool. I reproduced the clamp error, kept both components through a real view, and checked the real path against the previous code.