Replace floorf()-based implementation with direct IEEE 754 bit
manipulation. The optimized version avoids expensive function calls
and floating-point operations by directly manipulating mantissa bits
to achieve rounding.
Key optimizations:
- Fast paths for special values (Inf/NaN, |x| >= 2^23, |x| < 0.5)
- Add 0.5 in mantissa position then truncate fractional bits
- Pure integer operations until final conversion
- faster than floorf()-based approach
The implementation is bit-exact and maintains IEEE 754 compliance.