Release 0.8.4

SOLVAX 0.8.4 extends the implicit-differentiation and fixed-point layers:

• linear_solve accepts an independent transpose_solver and optional has_aux diagnostics, so the primal and adjoint solves can use different methods (or the same method with different tolerances) while preserving the implicit JVP and VJP behavior. See Matrix-free nonlinear solves and implicit differentiation.

• anderson_weights exposes the safeguarded Anderson weight solve as a standalone primitive, so one history’s weights can be reused across differently shaped coupled-state blocks. See Fixed-point acceleration.