[Feature Request] Parallelization of the CPU DequantizeLinear

[VamsiRatnakaramWork]

Describe the feature request

DequantizeLinear's current implementation is naive - single thread and scalar instructions.

#20901

Could we prioritize a MT/vectorized implementation for this code path to match the MlasQuantizeLinearKernel implementation?

@fajin-corp already made some comments about this.

https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/cpu/quantization/quantize_linear.cc#L302

Describe scenario use case

We should see performance improvements to a multitude of use cases.

Recently, the QNN-EP made this code path the default for the execution provider for performance reasons as well. It's likely vectorization would help this effort even more.

https://github.com/microsoft/onnxruntime/releases/tag/v1.20.2

Another user recently commented also about performance gains with Qwen 2.5 0.5B model:

#23395

The thread has since become stale, so I cannot add onto it.

[jywu-msft]

#24818

[jywu-msft]

initially this is only enabled when ORT_CLIENT_PACKAGE_BUILD build option is enabled to avoid unexpected perf impacts