autodiff vs hand diff for (I-A)⁻¹

[JeffFessler]

currently the optimizer for fitting bssfp signals uses autodiff.
i wonder if exploiting the properties of the matrix exponential and inverse could be more efficient?

[JeffFessler]

see also
JuliaDiff/ForwardDiff.jl#174