TensorCheck matmul skips rank validation: rank-1 matmul panics with `attempt to subtract with overflow` on ndarray

[ax1s-x1zz]

TensorCheck::matmul short-circuits for D < 2 in crates/burn-tensor/src/tensor/api/check.rs:

check = check.binary_ops_device("Matmul", &lhs.device(), &rhs.device());

if D < 2 {
    return check;
}

Tensor::matmul is defined generically over Tensor<D, K> where K: Numeric, so Tensor<1>::matmul compiles but skips all validation. Rank-1 inputs have no matrix dimensions, so the operation is ill-defined; the matmul docs only describe [..., M, K] @ [..., K, N] shapes.

What happens today, per backend:

• ndarray: panics in crates/burn-ndarray/src/ops/matmul.rs with attempt to subtract with overflow (it indexes shape_lhs[ndims - 2] with ndims == 1).
• tch: libtorch treats 1D x 1D matmul as a dot product and returns a 0-dim scalar, which does not match the Tensor<1> output rank.
• cubecl: not a valid kernel input.

None of these produce a usable Tensor<1> result; the backends fail in inconsistent, hard-to-debug ways.

Proposal: have TensorCheck::matmul reject ranks < 2 with a clear message (mirroring the TensorCheck::tri pattern) and add a should_panic regression test in burn-backend-tests/tests/tensor/float/ops/matmul.rs.