TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation