Separate plots

[cdhinrichs]

Addresses #292

This PR does the following:

• PRESERVES all pre-existing behavior, especially with respect to plotting
• Defines a new set of constants:
  • one for each type of plot that can be generated
  • a list of ALL plots that can be generated
  • lists of plots that are generated together
• Allows the plot argument to WeightWatcher.analyze() and other entry points to take either a boolean argument, which is replaced with the list of ALL plots if True, else an empty list; or to take a list of valid plot constants from the list above
• Adds a check to valid_params() to ensure the plot argument only has allowed constants in it
• Gates the creation of each separate plot on whether its corresponding constant is found in the params[PLOT] list or plot argument, whichever argument is passed to the function where plotting takes place.
• Adds a set of tests to ensure the above behavior
• Fixes a bug where MPDENSITY plots were not being saved to safedir.
• Fixes an error in test_torch_linalg relating to half precision conversion (float16).

The following tests are currently passing, or else have the same failures as in the master branch:

• Test_ValidParams
• Test_KerasLayers
• Test_PyTorchLayers
• Test_PyStateDictLayers
• Test_VGG11_noModel
• Test_Albert
• Test_Keras
• Test_ResNet
• Test_ResNet_Models
• Test_RMT_Util
• Test_Vector_Metrics
• Test_PyTorchSVD
• Test_PowerLaw
• Test_Plots
• Test_Pandas

The following tests have been skipped due to runtime:

• Test_VGG11_Distances
• Test_VGG11_Base
• Test_VGG11_StateDict
• Test_VGG11_Alpha_w_PowerLawFit

[cdhinrichs]

The bug was that there was a clause for overlaying the pdf plot on top of the log-log ESD. When I examined the code, it looked like that code was intended to run unconditionally whenever anything gets plotted, because it was interwoven with the clause for computing min_evals_to_plot, which does need to be done unconditionally whenever anything is plotted. The fix is to move the pdf overlay into the WW_PLOT_LOGLOG_ESD condition block. I plotted all of the singletons and verified that they all look as they should.