test_netcvode.py: increase comparison tolerances.

[olupton]

• This makes the test, which was introduced in Increase coverage of src/nrncvode/netcvode.cpp #1752, pass (locally) when NEURON is compiled with the NVIDIA compilers, nvc and nvc++.
• Using math.isclose for the comparison reduces the number of h.Vector objects that are constructed; modify the reference file to take that into account.
• The CI of Add missing utility header for std::exchange BlueBrain/CoreNeuron#818 is now testing this branch of NEURON with the NVIDIA compilers.

[nrnhines]

The question of whether it's best to use absolute or relative tolerance is hard to resolve. I think generally we are looking for accuracy over the dynamic range of each state. I.e. a channel gating state range is 0-1 and errors near 0, or .5, or 1 are in my opinion equivalent. Voltage generally ranges from -100 to 100mV and there is nothing special about 0. Concentrations range from >0 to 100mM but calcium is often special and we want it to be accurate in the 10-6 mM (nanomolar) range. I've written pump models in which the dynamic range was on the order of >0 to 1e-15 . So with CVode I'm often selecting atolscale for each state type based on the maximum absolute value during a sim for each state type (rounded to nearest order of magnitude). The the atol multiplier gives an overall indication of the accuracy desired in the sense that if atol is reduced by a factor of 10, the overall accuracy of the entire sim increases by a factor of 10. For testing, it's delightful if the results are identical . Otherwise, I guess relative tolerance is the best way to compare.

Seems like the nvidia change to number of time steps needs to be investigated but I'm happy to have that put into a separate issue.