Bugfix: wrong attention mask calculation resulted in wrong embeddings

[maziyarpanahi]

This pull request includes several changes to the BGEEmbeddings class in python/sparknlp/annotator/embeddings/bge_embeddings.py and various Scala files to add new functionality and improve code quality. The most important changes include adding support for the CLS token in sentence embeddings, updating the default pretrained model, and various code refactorings to improve readability and maintainability.

Enhancements to BGEEmbeddings:

• Added HasClsTokenProperties to BGEEmbeddings to support the use of the CLS token for sentence embeddings. (python/sparknlp/annotator/embeddings/bge_embeddings.py) [1] [2] [3]
• Updated the default pretrained model name to bge_small_en_v1.5. (python/sparknlp/annotator/embeddings/bge_embeddings.py)

New Class and Methods:

• Introduced HasClsTokenProperties class to handle CLS token properties. (python/sparknlp/common/properties.py)

Refactorings and Improvements:

• Refactored BGE class to include useCLSToken parameter in various methods and improved tensor handling. (src/main/scala/com/johnsnowlabs/ml/ai/BGE.scala) [1] [2] [3] [4] [5] [6] [7]
• Removed unnecessary blank lines and improved code formatting in Albert and Bart classes. (src/main/scala/com/johnsnowlabs/ml/ai/Albert.scala, src/main/scala/com/johnsnowlabs/ml/ai/Bart.scala) [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14]

These changes collectively enhance the functionality of the BGEEmbeddings class and improve the overall code quality in related files.This pull request includes several updates to the attention mask logic across multiple classes in the com.johnsnowlabs.ml.ai package. The changes primarily involve modifying the condition used to set the attention mask values.

Updates to attention mask logic:

• BGE.scala: Changed the condition for setting attention mask values from x < 0L to x == 0L in both getSentenceEmbeddingFromOv and getSentenceEmbeddingFromOnnx methods. [1] [2]
• E5.scala: Updated the attention mask condition from x < 0L to x == 0L in the getSentenceEmbeddingFromOnnx method. [1] [2]
• MPNet.scala: Modified the attention mask condition from x < this.paddingTokenId to x == this.paddingTokenId in the getSentenceEmbeddingFromOv and getSentenceEmbeddingFromOnnx methods. [1] [2]
• Mxbai.scala: Changed the attention mask condition from x < 0L to x == 0L in the getSentenceEmbeddingFromOnnx method.
• Nomic.scala: Updated the attention mask condition from x < 0L to x == 0L in the getSentenceEmbeddingFromOnnx method. [1] [2]
• SnowFlake.scala: Modified the attention mask condition from x < 0L to x == 0L in both getSentenceEmbeddingFromOv and getSentenceEmbeddingFromOnnx methods. [1] [2]
• UAE.scala: Changed the attention mask condition from x < 0L to x == 0L in both getSentenceEmbeddingFromOpenvino and getSentenceEmbeddingFromOnnx methods. [1] [2] [3]

[ahmedlone127]

LGTM!