Tflite micro python package uploaded to PyPI should be built with O2 optimizations

[pmjandreen]

Hi!

First of all I want to say thanks for the excellent tflite micro python package. The only issue I have encountered this far is its inference speed, very impractical to use for evaluations involving thousands of images. I tried building the package myself, editing tflite-micro/python/tflite_micro/pypi_build.sh, simply adding --copt=-O2 to the bazel build call (although I had to suppress a warning also to make it work). With the resulting package my particular model inference was over 4 times faster. Do you think building with optimizations could be made standard for the python packages uploaded to PyPI?

Best Regards,
Mikael