Optimize and fix GNNLSH

[jpata]

• GNNLSH attention kernels (Gaussian and Attention) and implemented inter-bin
• Expanded GNNLSHConfig with parameters for kernel_type, use_interbin_attention, and head dimensions.
• Fixed GNNLSH inter-bin attention export by using symbolic shapes and a manual attention implementation
• Added new test cases for GNNLSH attention variants in test_mlpf_models.py and introduced test_simple_mha_masking.py to verify attention masking and ONNX compatibility.
• Removed obsolete CMS data preparation scripts and Tallinn-specific file lists (zee_nopu.txt, zmm.txt, etc.).
• Updated LitePT tests to gracefully skip when dependencies are unavailable and applied general formatting fixes.

Compare the validation loss of transformer vs. GNNLSH:

Compare the runtime performance on A100 vs loss and ONNX export MAE: