better validation ux for attention mechanism selection

[bghira]

This pull request introduces improved validation for attention mechanism selection in both configuration normalization and server-side validation logic. The main focus is to ensure users receive clear error messages when selecting unsupported attention mechanisms, specifically addressing compatibility issues with xformers on certain GPUs and the availability of sageattention.

Attention mechanism validation improvements:

• Added is_sageattention_available and xformers_compute_capability_error helper functions to attention_backend.py to check for SageAttention installation and xformers GPU compatibility, respectively.
• Updated _normalize_input_args in cmd_args.py to raise a ValueError if xformers is selected on unsupported GPUs, or if sageattention is requested but not installed.
• Enhanced server-side validation in validation_service.py to add configuration errors for unsupported xformers or missing sageattention, improving feedback to users during config validation.

Internal API changes:

• Imported the new attention mechanism validation helpers into both cmd_args.py and validation_service.py to enable their use in argument normalization and validation logic. [1] [2]