Generalize train cloud microphysics

[rouson]

This PR adds a tensor_names_t object to the demo/training_configuration.json file and adjusts the train-cloud-microphysics demonstration application to read the object and use it to determine what training-data input and output variables to read from the NetCDF files produced by the neural-network branch of the Berkeley Lab fork of ICAR. With NetCDF files demo/training_input.f90 and demo/training_output.f90, a sample training run should produce output similar to the following:

cd demo
% ./build/run-fpm.sh run train-cloud-microphysics --flag -O3 -- --base training --epochs 10 --start 720 --bins 4
Project is up to date
 Reading physics-based model inputs from training_input.nc
 - reading potential_temperature
 - reading qv
 - reading qc
 - reading qr
 - reading qs
 - reading pressure
 - reading temperature
 - reading time
 Reading physics-based model outputs from training_output.nc
 - reading potential_temperature
 - reading qv
 - reading qc
 - reading qr
 - reading qs
 - reading time
 Calculating desired neural-network model outputs
 - dpotential_temperature/dt
 - dqv/dt
 - dqc/dt
 - dqr/dt
 - dqs/dt
 Defining input tensors for time step 720 through 720 with strides of 1
 Defining output tensors for time step 720 through 720 with strides of 1
 Defining a new network from training_configuration_t and tensor_map_t objects
 Conditionally sampling for a flat distribution of output values
 Keeping 39 out of 1250565 input/output pairs in 39 out of 1024 bins.
 Normalizing the remaining input and output tensors
 Training network
        Epoch  Cost (avg)
1    .14993742
2    .80107644E-1
3    .7153832E-1
4    .70149295E-1
5    .7109204E-1
6    .7106401E-1
7    .7154923E-1
8    .7048027E-1
9    .69384895E-1
10    .7029437E-1
 Training time:  2.6504E-02 for 10 epochs
 System clock time:  384.031288