feat: fundamental codebase refactor

[b4rtaz]

This PR introduces fundamental changes in Distributed Llama. Work in progress.

List of changes:

• ✅ Introduced an abstract neural network model with opcodes to describe neural network behavior
• ✅ Completely restructure the project
• ✅ Batch processing to support evaluation and prediction #138
• ✅ Speed up matmul operation for the evaluation (used sgemm llamafile)
• ✅ Improve tokenizer
• ✅ Fixed obvious memory leaks (detected by -fsanitize=address)
• ✅ ARM and AVX2 optimalisation for all op codes

Fixes:

• Fixed a bug in rope scaling
• Fixed a bug in the tokenizer that caused invalidly tokenized special tokens