(1) Fix an issue such that the augmentation may draw out-of-intended-boundary noise for concentration augmentation.

(2) Add optional least square solver (torch.linalg.lstsq) for concentration computation (concentration_method='ls', besides the existing 'ista' and 'cd' options) when dealing with huge number of small image inputs (e.g., batches of Nx3x256x256 input) on-the-fly. Note that while 'ls' is faster in such scenario, it does not have sparse constraints, and it may fail on GPU if the height/width of image is too large regardless of batch size, due to the limitation of torch.linalg.lstsq.

(3) Readme and demo update.