Add transformer model support, fix Multi-Head Attention quantization and enhance the pipeline

[federicobrancasi]

Add transformer model support, fix Multi-Head Attention quantization and enhance the pipeline

Summary

• Added comprehensive support for transformer-based architectures (CCT, ViT) to DeepQuant.
• Fixed critical Multi-Head Attention (MHA) quantization issues that prevented proper handling of packed projections and batch-first layouts.
• Enhanced the quantization pipeline with new utilities and improved error handling.

---

Key Changes

1. New models support

• Implemented Compact Convolutional Transformer (CCT) architecture
• Added Vision Transformer (ViT-B/32) support
• Created comprehensive test suites: TestCCT.py, TestCCTPretrained.py, TestVitB32.py, TestVitB32Pretrained.py
• Added ResNet-50 pretrained test to validate pipeline scalability on larger networks

2. Multi-Head Attention fixes (needed to make ViT's QuantMultiheadAttention work)

• Fixed MHA handling in CustomForwards/MultiHeadAttention.py to properly support:
  • Packed in_proj weights (common in transformers)
  • Batch-first tensor layouts with correct transposition
  • View operations that previously caused dimension mismatches
• Added support for self-attention scenarios where Q, K, V use the same input

3. Quantization improvements

• Changed rounding policy from round to floor in quantization modules for more accurate integer representation
• Added checkEquivalence flag to brevitasToTrueQuant() function to automatically validate quantized models against original outputs

4. ONNX Runtime enhancements

• Updated to ONNX opset version 18 for proper LayerNormalization folding required by Deeploy
• Implemented deterministic session handling by creating a function to disable all optimizations, ensuring exact reproducibility for Deeploy verification

5. Infrastructure improvements

• Fixed .view(-1) operations in tensor recording that caused shape inference issues
• Included CIFAR-10 dataset for CI testing of CCT models