ToT: strided arena kernels for complex inner cells; complex ToT scaling

[kshitij-05]

Problem

ToT×ToT contractions with std::complex<double> inner cells never use the strided arena kernels. arena_strided_dgemm_ce_e / _ce_ce_right / _ce_ce_left and their install gates in ContEngine are hard-coded to double, so a complex contraction falls back to the generic per-cell path: one tiny inner op per result cell and contracted index, instead of one strided BLAS gemm per cell. The kernels themselves are only layout/stride logic on top of blas::gemm, which already has the complex overloads.

Separately, scaling a complex ToT by an integral factor (2 * A("i;j")) does not compile: the ToT fill lambdas in Tensor::scale/add/subt/mult and ArenaTensor::scale_to use the raw elem * factor and miss the mixed complex×scalar operator* in detail that the non-nested scale branch already uses.

Changes

• The three strided kernels are templated on the inner numeric type; detail::is_strided_dgemm_numeric_v<T> admits float, double, complex<float>, complex<double>. All three operands must share the type.
• ContEngine install gates and the hc+e reuse gate use the trait; the diagnostic helpers' cell pointers are made type-generic.
• ToT scale/add/subt/mult fill lambdas and ArenaTensor::scale_to pick up detail's mixed operators.
• Tests: complex cases for the three kernels; a complex<double> row in the ToT fixture, so tot_expressions / tot_dist_array now cover complex ToTs.

Testing

arena_strided_dgemm, einsum*, tot_expressions, tot_dist_array_part{1,2}: 807 cases, no errors. With the change a complex ToT×ToT contraction reports strided ce+e installs where before it reported none.

Not addressed

• conj() on a ToT contraction does not compile (ComplexConjugate<void> has no conversion to the element type; ContractReduce<…, ComplexConjugate<void>> needs a value-returning gemm).
• The ToT × plain-tensor scale GEMM path requires identical element types; a complex ToT scaled by a real plain tensor uses per-cell AXPY.