fix signed overflow in fixed-point sample rounding before saturation

[netliomax25-code]

1. In the fixed-point build, output_to_PCM (libfaad/output.c) and real_to_int16 (libfaad/lt_predict.c) add a rounding bias to the raw int32 sample before the saturation clamp, so a decoded sample near INT32_MAX/INT32_MIN overflows the signed addition and the wrapped value then skips the clamp, emitting a garbage sample instead of the saturated one. UBSan on an HE-AAC stream: signed-integer-overflow at output.c:491 (2147483384 + 8192) and lt_predict.c:140.
2. The same rounding-before-saturation pattern is present in the 16-, 24- and 32-bit output_to_PCM cases and in real_to_int16.
3. Apply the rounding bias only when it cannot overflow the int32; a value large enough to overflow saturates anyway, so valid output is unchanged. Verified byte-identical 16/24/32-bit output on LC, HE-AAC and HE-AACv2 streams, and the CMake build stays clean with -Werror=strict-aliasing. Float build is unaffected.