v0.7.1 speedup, for dev only

For dev only

Speed up by:

• default n_iters = 101 now
• removing unnecessary and duplicated calculations
• in place calculations whenever possible
• no_grad when no need
• better shape initialization and tensor slicing to have fewer reshape calls
• no large memory copy between CPU and GPU inside the optimization loop

Not yet put to the code base:

• remove calculate_metrics (21-sec to 17-sec speed up for the following test case), need to retain the inside percentage
• sample coords to mimic stochastic gradient descent (the random indices generation might take time, might reduce accuracy and then need more n_iters)
• separate molecules so as to remove the div by num_atoms averaging calculation (this might also allow better memory usage)

(.venv) D:\GIT\DiffFitStar\demo\5wvi>python D:\GIT\DiffFit\src\DiffAtomComp.py --target_vol D:\GIT\DiffFitStar\demo\5wvi\6693_seg\emd_6693_region_4208.mrc  --target_surface_threshold 0.3  --structures_dir D:\GIT\DiffFitStar\demo\5wvi\AF_input --out_dir D:\GIT\DiffFitStar\demo\5wvi\DF_torch_profile --out_dir_exist_ok True --N_shifts 10 --device cuda --n_iters 101

Full Changelog: v0.7.0...v0.7.1