- split sample programs: count_mac, debug_model, sample_xxx
- added clang-format file

Improvements:
- improve SIMD support for float (AVX512, conv1x1 etc.)
- go back to HWCD loop order. Some gains on big models.
- nbOutputs() method in Model

Corrections:
- init is now re-entrant: it is possible to do inference with multiple input sizes by calling init with the new size
- remove all MAC counters in SIMD paths + add warning message

Cleaning:
- clean the naming for conv2d (simd, dispatch etc.)
- clean messages for SIMD support
- clean macro for debug
- clean sample_test.sh script
- clang format everything
- copyright year
- README update
- other minor cleanings