refactor: refinements in utils and associated functions

[neuralsorcerer]

Changes:

• Replaced activation and loss dictionaries with Enum classes ActivationType and LossType for type safety.
• Replaced the old initialization helper with a clearer xavier_init_weights function for linear layers.
• Neural linear regression selects its output activation via ActivationType for clarity.
• Rewrote the normalized Softplus activation to avoid redundant computations and added a default Softmax dimension for the activation map.
• Fixed the convolution block to apply batch normalization using the correct output channel size.
• Neural bandit classes accept LossType parameters and call loss_type.function() when training.

Why these changes?

• Mostly mentioned in TODOs in utils.py i.e. (Enums and xavier_init_weights), so decided to refactor according to it.
• MLP construction now instantiates activations through the enums, reducing string lookups.

[facebook-github-bot]

@rodrigodesalvobraz has imported this pull request. If you are a Meta employee, you can view this in D78998902.

[facebook-github-bot]

@ayushj240 merged this pull request in 9de0995.