Unveiling Project-Specific Bias in Neural Code Models