Fix for latest Arm Compiler

TensorFlow Lite Micro build produced warnings/errors with the latest
Arm Compiler (version 6.22) because of conflict between floating
point standard conformance and optimisation level.

As the TensorFlow Lite Micro's Makefile uses `-ffp-mode=full` for the
Arm Compiler, the fix is to use `-O3` for compiling the kernels. For
CMSIS DSP library we set -ffp-mode=full` for all Arm Compilers.

The issue was first raised here:
https://discuss.mlplatform.org/t/tensorflow-build-failure-with-arm-compiler-6-22/287



Change-Id: I5c19c4016a922b596b84937a366d261fd9570090
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>