feat: add Dockerfile for containerized fine-tuning

[abdelhadi703]

Summary

Adds a Dockerfile for containerized fine-tuning, addressing recurring environment setup issues (#98, #109, #92).

What's included

• Dockerfile: Based on pytorch/pytorch:2.2.0-cuda12.1-cudnn8-devel with all dependencies pre-installed
• .dockerignore: Excludes unnecessary files from the build context
• README update: Docker usage instructions (single-GPU and multi-GPU)

Usage

# Build
docker build -t mistral-finetune .

# Single GPU
docker run --gpus all -v /data:/data -v /model:/model mistral-finetune --config /data/config.yaml

# Multi-GPU (torchrun)
docker run --gpus all --entrypoint torchrun mistral-finetune \
  --nproc_per_node=4 /app/train.py --config /data/config.yaml

Design decisions

• Base image: pytorch/pytorch:2.2.0-cuda12.1-cudnn8-devel matches the pinned torch==2.2 requirement and includes CUDA development headers needed for xformers compilation
• Volume mounts: Training data and model weights are mounted at runtime (not baked into the image) for flexibility
• Entrypoint: Defaults to python -m train for simplicity; override with torchrun for distributed training

Related issues

• Partially addresses [BUG: validate_data.py ModuleNotFoundError (finetune & tensorflow) #98 (environment setup issues)
• Partially addresses [BUG: No operator found for memory_efficient_attention_forward with inputs #109 (installation difficulties)