check disp_format and clamp if necessary

[jsalotti]

Problems

• when creating a Disparity object from a prediction, the tensor contains both positive and negative values. This causes a problem when converting between signed and unsigned disparity format
• when converting Disparity to Depth, some infinite values can happen in the resulting tensor

Fixes

• when creating disparity with "signed" format : force user to give camera_side and clip values to have only one sign (positive for right camera, negative for left)
• when creating disparity with "unsigned" format : clip values to have only positive
• when converting to depth : add a max_depth parameter to avoid infinite values