[FEATURE] Analytical gradients similar to how NONMEM does mu-referencing

[cbdavis33]

Description

Section I.44 in this NONMEM User's Guide and this NONMEM tutorial describe NONMEM's MU-referencing. In a NONMEM vs. Pumas vs. Stan/Torsten comparison for ACoP 2023, NONMEM's ESS/sec was substantially higher (even after all possible parallelizations and optimized efficiency), particularly for simple (analytical/mat. exp.) models, and we believe it is due to the analytical gradient rather than autodiff that NONMEM uses.

Example

$pmx\_solve\_1cpt(\dots)$

Expected Output

I expect that nothing from the user's perspective would change. pmx_solve_(two)onecpt(), pmx_solve_linode(), and pmx_solve_rk45(bdf, adams) should return the same as now.

Current Version:

v0.90.1