Fix wrapping handline

Several fixes to correctly handle memory wrapping:

• Bad in-place CP model for Conv and Pool (not enough margin, did not work with OpenMP);
• I/O offsets (usually named iOffset, oOffset) must stay signed, otherwise memory wrapping won't work (@pineapple);
• INPUT_MEM_STRIDE was used instead of INPUT_MEM_CONT_SIZE at some places;
• Wrapping was not handled by outputs save and cmp (@axelfarr);
• Fixed several signed/unsigned comparison warnings;
• Use size_t instead of int when possible.

This MR fixes sporadic failure of CI due to wrong computation in lenet.py with --mem_wrap argument.