Refactor: Improve modularity and readability across codebase

[Mirza-Samad-Ahmed-Baig]

This commit introduces several refactorings across run.c, test.c, and
train.py to enhance code modularity, readability, and maintainability.
The primary goal was to reduce function complexity and nesting depth by
extracting distinct logical blocks into dedicated helper functions.

Key changes include:

run.c:

• Extracted multihead_attention function: The complex multihead
  attention logic within the forward function has been moved into a
  new, self-contained multihead_attention function. This significantly
  reduces the nesting level and improves the clarity of the main
  forward loop.
• Extracted process_utf8_bytes function: The intricate UTF-8 byte
  processing within the encode function was isolated into
  process_utf8_bytes. This simplifies the encode function and makes
  the byte-level operations more understandable.
• Extracted render_chat_prompt function: The logic responsible for
  formatting chat prompts in the chat function has been moved to a new
  render_chat_prompt helper, making the chat function's flow clearer.
• Extracted get_chat_prompts function: The logic for acquiring
  system and user prompts within the chat function has been
  encapsulated in get_chat_prompts, further streamlining the chat
  function.

test.c:

• Extracted individual test_prompt_encoding functions: The
  monolithic test_prompt_encodings function was broken down into
  smaller, more focused functions (e.g., test_prompt_encoding_0,
  test_prompt_encoding_1, etc.). This improves test readability and
  makes it easier to identify and debug specific test failures.

train.py:

• Extracted setup_ddp function: The distributed data parallel (DDP)
  setup logic has been moved into a dedicated setup_ddp function,
  centralizing configuration and improving the main script's clarity.
• Extracted initialize_model function: The model initialization
  logic, including loading from scratch or resuming from a checkpoint,
  is now handled by initialize_model.
• Extracted setup_optimizer function: The optimizer setup, including
  GradScaler initialization and loading optimizer state from
  checkpoints, has been moved to setup_optimizer.
• Extracted save_checkpoint function: The logic for saving training
  checkpoints has been encapsulated in a save_checkpoint function,
  promoting reusability and cleaner code within the training loop.

These changes collectively contribute to a more organized,
readable, and maintainable codebase, making it easier to
understand, debug, and extend in the future.