[QNN EP] Add LowPowerBlockQuantization support for Gemm node

[quic-tirupath]

Description

• Low Power Block Quantization(LPBQ) is widely used to accelerate accuracy sensitive models via QNN(Qualcomm Neural Network) stack.
• LPBQ encoding format is Qualcomm's alternative for BlockQuantization technique.
• The current implementation expects LPBQ encodings packed in a node sequence (DQ -> Q -> DQ)
• This PR folds LPBQ pattern on Weight of Gemm nodes into a Qnn BlockExpansion encoding structure.
• This PR adds INT4 Quantization support

Motivation and Context

• This enables acceleration of accuracy sensitive models that require block quantization kind of encodings via QNN EP
• This avoids fallback of nodes consuming block quantized tensors on CPU EP and further improves inference time.

[Copilot]

Pull Request Overview

This PR adds Low Power Block Quantization (LPBQ) support for Gemm nodes in the QNN (Qualcomm Neural Network) execution provider. LPBQ is an alternative block quantization technique that enables acceleration of accuracy-sensitive models by avoiding CPU fallback for block-quantized tensors.

• Introduces a new fusion pattern to detect DQ->Q->DQ sequences on Gemm weights and convert them to QNN's BlockExpansion encoding
• Adds INT4 quantization support through specialized template traits and quantization functions
• Extends the quantization parameter wrapper to handle LPBQ encodings with per-channel float scales and per-block integer scales

Reviewed Changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
qnn_utils.h	Adds LPBQ data quantization function and Int4 quantization traits
qnn_utils.cc	Implements LowPowerBlockQuantizeData function for LPBQ encoding
qnn_quant_params_wrapper.h	Extends wrapper to support LPBQ quantization parameters
qnn_quant_params_wrapper.cc	Implements LPBQ constructor and deep copy logic
utils.h/cc	Adds utility functions for parent/child node traversal in fusion detection
qnn_node_group.cc	Registers LPBQ Gemm fusion and updates fusion dispatch logic
lpbqgemm_fusion.h/cc	Implements the LPBQ Gemm fusion pattern detection and QNN node creation
qnn_model_wrapper.h/cc	Templated UnpackScales function to support both float and uint8_t scales

[jywu-msft]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows x64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[jywu-msft]

There are build errors in both Linux QNN and Windows ARM64 QNN CI pipelines

[HectorSVC]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows x64 QNN CI Pipeline, Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[HectorSVC]