general product: handle fused-broadcast-on-the-right (rank-0 right operand)

[evaleev]

Problem

A general product whose right operand is entirely fused with no contraction — a fused broadcast on the right, e.g. C("b,k") = A("b,k") * B("b") — folds to a rank-0 right operand. TiledArray cannot host that: rank-0 Ranges are null (volume 0) by design, and gemm/GemmHelper assert operand.rank() == gemm_helper.rank().

The general-product engine already avoids rank-0 folded operands with a synthetic unit mode carried only in the GemmHelper, but ContEngine::synthetic_unit_left_external() covered only two of the three degeneracies:

• rank-0 result (no-external product), and
• rank-0 left (fused broadcast on the left).

The rank-0 right case was unhandled — a left-external unit can't fix it (right_rank stays 0). So folding the right operand aborted in BatchedContractReduce::contract_pair (Tensor::reshape volume assert), or silently mis-shaped the tile with asserts off.

This surfaced via MPQC PNO/CSV-CCSD denest products (A(b,k;…)·B(b;…) with b a per-pair batch index) evaluated through the native ToT*ToT->T dispatch added in #567.

Fix

Add a symmetric synthetic unit right-external mode, mirroring the existing left one:

• cont_engine.h: new synthetic_unit_right_external() (detects outer_size(right_indices_) == n_fused_modes_); u_right threaded through init_struct_general (both op_ ctors; right_op is NoTranspose for the synthetic mode), make_trange_general and init_distribution_general (neB -= u_right).
• sparse_shape.h: SparseShape::gemm_batched (shape-level analog) detects the right phantom by the one-rank mismatch and guards the right-outer loops / result rank; its fold_range lambda gains an append_unit option.
• batched_contract_reduce.h: contract_pair detects unit_right_external, excludes it from neB, and pads the folded right (and accumulating result) views with a trailing unit; the member fold_range gains an append_unit option.

The synthetic mode lives only in the GemmHelper; tranges, shapes and tiles carry the true ranks. The result gains a trailing unit mode (squeezed out of the actual result) exactly as the left case prepends one.

Tests

• general_product: new dense, ToT*ToT->ToT, and block-sparse fused-broadcast-right cases, all through the expression layer (which respects operand order — unlike einsum, whose operand reordering hides the shape).
• dot_inner: new ->T denest fused-broadcast-right case.

All general_product (61) and dot_inner suites pass, plus einsum_tiledarray/einsum_tot/einsum_tot_t/arena_einsum_unit_suite/expressions_suite (112) with no regressions. Also verified end-to-end against the originating MPQC CSV-CCSD case (correct to ~1e-11; previously crashed under asserts / silently wrong under Apple Accelerate).