perf: split the harmonic coefficient matrix into its independent blocks

[oameye]

rearrange! solves the linear system relating the harmonic equations to the derivatives of the harmonic variables. For an n-harmonic ansatz that matrix is n uncoupled 2x2 blocks, but it was eliminated whole. A Bareiss entry carries the determinant of everything eliminated so far, so by pivot k the polynomial arithmetic runs on k blocks' worth of determinant, and every solution comes out over the determinant of the full matrix instead of the single block determinant that survives.

fraction_free_linear_solve now finds the connected components of the bipartite row-column graph (union-find over the structural nonzeros) and eliminates each on its own. An unbalanced component is structurally singular and is handed to the full solve to reject, so nothing that used to throw now silently succeeds.

Over six representative systems rearrange_standard drops from 134.2 MiB to 10.1 MiB and from 0.114 s to 0.028 s. The worst case, two coupled oscillators at two harmonics, goes from 80.8 MiB to 3.1 MiB.

The solutions are deliberately not byte-identical to before: each now carries only its own block's determinant, so it is smaller. Verified by substituting them back into the source equations at random numeric points; the largest residual across all six systems is 9.4e-15.

Alongside this:

• Sums and products are rebuilt with SymbolicUtils' n-ary add_worker/mul_worker rather than folded with sum/prod/+=, which is quadratic in the number of terms. A solved Bareiss entry has thousands of monomials.
• simplify_exp_products missed exp(a)*exp(a): a product stores a repeated factor as a power, so it arrives as the single factor exp(a)^2 and the isexp test skipped it.
• collapse_pythagorean committed to the first squared trig factor it found in a summand, leaving cos(3t)^2*sin(t)^2*a + sin(3t)^2*sin(t)^2*a uncollapsed. Every way of splitting a summand is now a candidate partner.
• is_rearranged asked whether a derivative appears on the left-hand side by rendering both sides to strings and calling occursin. It now walks the tree.
• trig_reduce called expand_all where expand suffices: the extra Postwalk(expand_exp_power) was the single most expensive step of the averaging, and the simplify_exp_products on the next line already normalises exp(a)^n at every node it reaches.
• declare_variable built its variable with @eval and bound the result inside QuestBase. Nothing read that binding, and evaluating into a closed module makes the package impossible to precompile with a workload.