full refactoring of conv2d dispatch, peeling and support for pad 0

• refactor (simplify) the dispatch of the possible conv2d
• simplify the peeling for conv2d ixj
• add support for padding 0 by adding a template offset on the output
• reduce memory usage (but forbid multi-thread at layer level for now)
• cleaning
• add a macro to trace function calls (in conv2d only for now)