Add -DFLT_EVAL_METHOD=0 and -ffp-contract=fast

This patch adds both FLT_EVAL_METHOD set to 0 (to avoid double_t
being evaluated as long double) and -ffp-contract=fast to clang
can use FMA on targets which supports it.

The -ffp-contract=fast is specially required for aarch64 and the
optimized implementations are targeting explicit architectures
with contracted floating-point operations.

Bug: N/A
Test: ran tests

Change-Id: I4719e909e84435f8017a3d5a6dca87040b9d8aa5