feat: scale a batched tensor by a vector of per-batch scalars

[wardjm]

Closes #12.

A plain numeric array multiplying or dividing a tensor is now read as batch-wise scalar weights: weights * ω, ω * weights and ω / weights all work. The weights' shape broadcasts against the tensor's batch axes with numpy's right-alignment (reusing compatible_batches / align_batch_pair), and the coefficient axis is never matched against them. Weights may also expand the batch shape, so an unbatched form times a length-B vector yields batch (B,) — the same rule as upstream's _broadcast_batch_scalars + _broadcast_coeffs_to_batch.

This makes a spectral filter work as it does in Python:

vals, vecs = spectrum(up_laplacian(dg, 1) + down_laplacian(dg, 1))
exp.(-vals) * vecs      # weight each eigenform by a function of its eigenvalue

Tests

The @test_broken in test/test_pyclasses.jl (product_operations → per-batch scalar vector scaling) is flipped to real assertions and extended to cover the reverse order, division, batch expansion, multi-dimensional batch broadcasting, and the error path when no batch axis can absorb the weights. The upstream test that motivated the issue, test_operators.py::test_spectrum_eigenvector_batch_scaling_with_numpy, is ported into the operators section.

Full suite passes: 2381 tests, none broken.

Two deliberate divergences

The dispatch signature is AbstractArray{<:Number}, not the AbstractVector the issue suggested. Multi-dimensional batch shapes need the wider signature and upstream's np.broadcast_shapes accepts any rank — but it does mean a matrix times a tensor is now a batch-broadcast scale rather than a MethodError, so Julia's usual linear-algebra reading of A * x doesn't apply here.

Incompatible weights raise AssertionError, where upstream raises TypeError. That matches how + and the pointwise products already guard in this file, so it is consistent with the repo rather than with Python.

Not covered

weights .* ω (broadcast syntax) still fails — AbstractTensor has no Base.broadcastable. Upstream reaches it via __array_ufunc__, which this test file already declares out of scope. Filed separately as #19.