Optimized onnx transform class via multithreading

[abhishek-singh591]

ONNX Transform Optimization with Thread Pool

Refactored the ONNX transform class to use a thread pool for parallelizing tensor operations, replacing the previous iterative loop. This resulted in a notable performance boost in the FP16Clip transform and a marginal improvement in split_tensor(), which may have further optimization potential.

Performance (LLaMA 3.1 8B)

Operation	Without Thread Pool	With Thread Pool
FP16 Transform	51.94 sec	38.78 sec
Split Tensor	15.30 sec	13.73 sec

Thread count is hardcoded to os.cpu_count() * 4 to better handle I/O-intensive workloads. Performance may vary depending on the machine's threading capabilities, so results may not be exactly reproducible across environments.

[vbaddi]

@abhishek-singh591 this is good start, Can we verify a larger model. Say 70B?
Can you check on this model card: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct

[abhishek-singh591]

@vbaddi

LLaMA 3.1 70B Performance Comparison

Without Thread Pooling

• Execution time for f16: 419.0391 seconds
• Execution time for Split: 194.3422 seconds

With Thread Pooling

• Execution time for f16: 302.9838 seconds
• Execution time for Split: 163.3579 seconds

[abhishek-singh591]

@vbaddi
Updated the llama 3.1 70B results, hope this suffices for the scalability check.