Add (IA)^3 implementation

[calpt]

This PR adds the (IA)^3 adapter architecture proposed in "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning" by Liu et al. (2022).

• Paper: https://arxiv.org/pdf/2205.05638.pdf
• Reference implementation: https://github.com/r-three/t-few
• Implementation in adapter-transformers:
  • IA3Config class that inherits from LoRAConfig
  • New config options attn_matrices, composition_mode and no_decomposition to support scaled composition of (IA)^3