pyCUDA version of transient F-stat

See merge request GregAshton/PyFstat!13

- Add transient window type
- Rename arrays -> coord_arrays to solve caching check bug

 -previously check was done in _optional_import
  [called from init_transient_fstat_map_features()]
 -this was the right place only back when still importing autoinit
 -now do check at the time of make_default_context()

 -make some functions internal in tcw_fstat_map_funcs.py
 -clean up debug messages
 -clean up code formatting

Conflicts:
	pyfstat/core.py

 -not sure why this is not inherited from the base ComputeFstat class

 -requires manual context initialisation and cleanup

TransientGridSearch: new F_mn writer function matching the format of lalpulsar.write_transientFstatMap_to_fp

 -optional import of pyCUDA package only if requested
 -including kernel .cu files in packaging

 -to later allow for CUDA implementation
 -some optional-import acrobatics
  to fail gracefully if a module (especially CUDA)
  is not available
 -as of this commit, nothing beyond lal implemented yet

writer function for atoms (TransientGridSearch)

See merge request GregAshton/PyFstat!14

Closes #12

displayless matplotlib import workaround also in helper_functions.py

See merge request GregAshton/PyFstat!12

transients: fix window=none case and allow custom dt0, dtau

See merge request GregAshton/PyFstat!11

This updates get_input_data_array. First, the list of possible search
parameters is defined as a list of strings in the class itself avoiding
repeated definitions of get_input_data_array. Second, is `args.clean == True`
then the input_data is not generated as an array, but a generator. For
large arrays this avoids memory issues.