[AhmadHassan-BTed]

Added Int4BlockCompressor in fedlab/utils/compressor.py to compress serialized model parameters and gradients down to packed 4-bit integers with per-block scaling.

Background:
SerializationTool converts model parameters and gradients into flat 1D float32 tensors, and ModelMaintainer.shape_list notes intended use for tensor compression. Transmitting uncompressed 32-bit floats over network rounds is bandwidth heavy in distributed federated settings.

Changes:

• Added Int4BlockCompressor in fedlab/utils/compressor.py with:
  • Block-level dynamic scaling (scale = max(|x|) / 7.0).
  • INT4 quantization with two 4-bit values packed into one uint8 byte (~8x payload size reduction).
  • Optional residual error feedback tracking to mitigate gradient drift across communication rounds.
  • Convenience methods compress_tensor, decompress_tensor, compress_model, deserialize_model, compress_model_gradients, and deserialize_model_gradients.
• Exported Int4BlockCompressor in fedlab/utils/init.py.
• Added unit tests in tests/test_utils/test_compressor.py covering shapes, padding, reconstruction MSE, error feedback accumulation, and model serialization pipelines.