v3.3.2: fixing notebook filtering, parse_pose_star optics groups, .txt inputs for write_star

This patch release makes some improvements to tools used in writing and parsing .star files, as well as addressing a few bugs that have recently come to our attention:

• the filtering notebook cryoDRGN_filtering was very slow to run when applied to experiments using --ind; we tracked this down to using an incorrect approach to loading the dataset (#374)
• nicer FSC plots in backproject_voxel using code refactored to apply the methods used in fsc and plot_fsc
• fixing an issue when the total particle count was module one the batch size, causing dimensionality errors with the final singleton batch due to how some torch and numpy operations handle singleton dimensions (#351)
• creating a stopgap for #346 while we figure out what upstream problems could be causing these issues with analyze
• adding linspace=True to the np.linspace operation in pc_traversal for completeness
• properly supporting .txt files for write_star, with the correct file names now being written to the output, as well as --ind working correctly
• adding support for RELION 3.1 input files with multiple optics groups in parse_pose_star

We have also consolidated and improved upon several aspects of our continuous integration testing setup, including new tests covering the cases described above, refactoring the data fixtures used in existing tests, and testing across multiple torch versions after finding issues specific to v1.8 in analyze_landscape_full.