No condition on text & load models from external directory

[peymvn]

I made two minor changes:

1. Added an option for condition_on_previous_text with the default value set to False. I noticed that Whisper often gets confused when conditioned on previous text, especially during scenes with unusual background sounds (e.g., crowd noise, overlapping speech, etc.), particularly when the language is not English. In these cases, even a quick moment of a confusing sound can generate nonsense that ruins the next 30 seconds.

2. Added model_weights_root to avoid re-downloading the large checkpoints. I already have them stored elsewhere, so this is convenient for my setup and people like me.

[what-the-diff]

PR Summary

• Inclusion of .vscode/ directory in .gitignore
  We've updated the .gitignore file to hide settings specific to the personal Visual Studio Code IDE, which helps maintain a clean codebase and avoid unnecessary version control changes.

• Introduction of new command-line arguments in cli.py

  • --model_weights_root: Allows users to specify the base directory for storing the model's weight files, affording flexibility and improved organization.
  • --condition_on_previous_text: Enables the model to condition its output on previous text, facilitating continuity and coherence in the model's output.

• Use of model_weights_root for model loading
  The PR modifies the line of code that loads our Whisper model, which now takes advantage of the model_weights_root parameter, reinforcing the use of specified directories for model weight storage.

• Update to subtitle retrieval process
  The subtitle retrieval process now involves the condition_on_previous_text parameter. This change helps to improve the relevance of transcriptions provided by our service, as they can now take into consideration prior textual context.