Respect conjugate transpose in complex vector mode products - #696

For complex vectors, mode_dot(..., transpose=True) ignores conjugation in both tensor-algebra implementations. The core multi_mode_dot already conjugates vectors, while its einsum implementation does not. Consequently, a single mode product and its one-factor multi-mode equivalent disagree.

For X=[[1+2j,3-1j],[-2+1j,4+3j]] and v=[1+1j,2-1j], the Hermitian mode-0 contraction is [-2+1j,7+6j]; both single-mode implementations currently return [-4+7j,15+4j].

Apply the existing Hermitian transpose convention to vector branches using the wrapped backend conj. Update the relevant parameter documentation and add one compact parameterized regression covering both vector modes, transpose values and complex64/128. Default bilinear calls and real-vector values retain their semantics.

Validation on exact main cdde62f:

• Exact-base final regressions: 4 failures and 4 controls for each tenalg implementation.
• Entire tenalg test directory: 367 passed in each NumPy/PyTorch × core/einsum configuration, with 16 existing SVD warnings per run. The final formatted regression also passes all four configurations.
• Separate scalar-loop complex contractions and independent review cover mixed matrix/vector products and real/imaginary finite-difference gradients. Black and diff checks pass.

CPU NumPy/PyTorch were tested; other backend and GPU matrices were not run. Repeated/negative modes and reordered skip behavior are outside this change. Implementation, testing and description were AI-assisted.