Add configurable --seed argument for reproducible training

[Vikasboura]

This PR adds a configurable --seed CLI argument to the training scripts in
Transformers_Classification_DeepLense_Kartik_Sachdev to enable reproducible
experiments.

Problem

Some training scripts either used a hardcoded seed or did not initialize
random seeds at all, which made experiment results difficult to reproduce.

Changes

• Added --seed CLI argument (default: 42) to:
  • main.py
  • simsiam.py
  • pretrain.py
  • finetune.py
  • finetune_byol.py
• Reused the existing seed_everything() utility from utils/util.py
  to initialize random, numpy, torch, and CUDA seeds.
• Updated README.md with a Reproducibility section and example usage.

Example

If --seed is not provided, the default behavior remains unchanged
(default seed = 42).