* decouple memory copy

* fix ci bugs

* fix ci compile bugs

* fix rocm compile

* fix ci bugs

* add unittest for einsum-v2-trace and diagonal

* repeat labels.

* einsum support repeated labels.

* forward is ok for diagonal and undiagonalized.
TODO: check backward is ok by our theorem.

* backward is ok!

* fix by PR suggestions.

* fix ci error

* fix ci error

* fix ci warning

* phi_fill_diagonal_tensor

* delete extra lines

* update

* add legacy api test

* rename sig