llama.cpp adds GGML_OP_LIGHTNING_INDEXER for DeepSeek V3.2/V4

The llama.cpp project has introduced a new `GGML_OP_LIGHTNING_INDEXER` operation in ggml to implement the lightning indexer used by DeepSeek V3.2 and V4 models.

• The implementation removes scale parameters from the operator and adds an f16 mask parameter.
• It includes fixes for lightning indexer mask broadcasting and checks that input tensors are not transposed.
• Tests were added to verify the new operation, counting FLOPs instead of bandwidth.
• The `llama` component now utilizes this new operator for DeepSeek V3.2 and DeepSeek V4 models.

This update enables efficient inference support for DeepSeek's latest architectures within the llama.cpp framework.