Develop a sufficiently large benchmark problem

This now reports for nonlinear-jacobian-apply on agamemnon:
samples: 442 (excluding 3 warmup samples)

host flavour:
mean: 0.0118471s
stddev: 0.000154351 (1.30285%)

device flavour (cuda):
mean: 0.00529035s
stddev: 9.87994e-05 (1.86754%)

This is far from being O(1second), but making the problem much larger would result in much more iterations of the linear solver, making the overall runtime of the program unbearable.

Since the standard deviation seems small compared to the mean runtime, accept this setup.

Closes: #62 (closed)