Fixed batch processing

[jkrukowski]

Key Changes

• Added attention mask support: Tokenizer now returns BatchTokenizeResult struct containing both tokens and attention masks
• Implemented masked mean pooling: Padding tokens are now correctly excluded from pooling calculations
• Updated all embedding models: Bert, CLIP, ModernBert, Roberta, and XLMRoberta now use attention masks in forward pass
• Refactored accuracy tests: Split monolithic test file into focused per-model test suites with batch accuracy tests
• Enhanced test infrastructure: Added shared utilities and updated Python generation script for batch testing

Models Updated

• BertModel
• ClipModel
• ModernBertModel
• RobertaModel
• XLMRobertaModel

Breaking Changes

• tokenizeTextsPaddingToLongest methods now return BatchTokenizeResult instead of [[Int32]]

co-author: @dang-hai