[QNN EP] Add use_native_matmul session config for native MatMul lowering

[qti-mbadnara]

Description

Adds an ep.QNN.use_native_matmul session config key (default "0") that routes MatMul and Gemm operators through QNN_OP_MAT_MUL instead of QNN_OP_FULLY_CONNECTED, eliminating the host-side weight transposition that occurs during session creation.

---

Motivation & Context

ORT QNN EP lowers MatMul (with a rank-2 static weight) and all Gemm operators to QNN_OP_FULLY_CONNECTED. FC requires the weight in [N, K] layout, but ONNX stores it in [K, N], so the builder pre-transposes every static weight at session creation time.

QNN_OP_MAT_MUL accepts weight in the natural ONNX [K, N] orientation, making it a drop-in replacement that avoids this transposition entirely. Compared to QNN_OP_FULLY_CONNECTED, it also supports a larger max rank (5 vs. 4), which removes the need to flatten leading activation dimensions for batched inputs.

Paths left on FC (not affected by this flag):

• Block-quantized (BW_FLOAT_BLOCK) Gemm — BQ weight layout and quantization params are FC-specific
• Absorbed-Reshape QDQ Gemm (output_reshape_node != nullptr) — narrow QDQ pattern, left for follow-up
• LPBQ MatMul / LPBQ Gemm fusions — unaffected; LPBQ MatMul already emits QNN_OP_MAT_MUL natively

---

Supported Configuration

Config key	Value	Behavior
ep.QNN.use_native_matmul	"0" (default)	MatMul and Gemm route to QNN_OP_FULLY_CONNECTED. Weight transposed to [N, K] at session creation.
ep.QNN.use_native_matmul	"1"	MatMul and Gemm route to QNN_OP_MAT_MUL. Weight stays in [K, N]; transpose_in0/transpose_in1 params set from transA/transB attributes.

---

Unsupported Configuration

• Block-quantized (BW_FLOAT_BLOCK) Gemm — always emits QNN_OP_FULLY_CONNECTED regardless of this flag; BQ weight encoding is FC-specific.
• Absorbed-Reshape QDQ Gemm (output_reshape_node != nullptr) — left on FC; targeted for follow-up.
• LPBQ MatMul / LPBQ Gemm fusions — independent node group fusions, unaffected.

---

Data Types

Both QNN_OP_FULLY_CONNECTED and QNN_OP_MAT_MUL share identical data type support. Switching between them has no effect on supported dtypes.

Backend	Supported dtypes
CPU	FP32
HTP	FP16, BF16, INT8, INT16 (Firewheel only), BW_FLOAT_BLOCK FP16+SFIXED8 (V69+)
GPU	FP16, FP32

---

Files Changed

File	Change
onnxruntime/core/providers/qnn/builder/qnn_model_wrapper.h	Added bool use_native_matmul = false to ModelSettings. Threaded to op builders via GetModelSettings().
onnxruntime/core/providers/qnn/qnn_execution_provider.cc	Parses ep.QNN.use_native_matmul via ParseBoolOption (default false).
onnxruntime/core/providers/qnn/builder/opbuilder/matmul_op_builder.cc	Guard in CheckInputs: forces use_fully_connected = false when flag is set. Existing QNN_OP_MAT_MUL path handles rank-2 static weights correctly without transposition.
onnxruntime/core/providers/qnn/builder/opbuilder/qlinear_matmul_op_builder.cc	Same guard in DecideUseFullyConnected.
onnxruntime/core/providers/qnn/builder/qnn_node_group/reshape_gemm_fusion.cc	Early return nullptr in TryFusion2/3/4 when flag is set, allowing nodes to fall through to GemmOpBuilder.
onnxruntime/core/providers/qnn/builder/opbuilder/gemm_op_builder.cc	(a) ProcessInputs: skips host-side weight transposition when flag is set. (b) ProcessAttributesAndOutputs: normal path and FC+Add decomposition path emit QNN_OP_MAT_MUL with transpose params derived from transA/transB.
docs/execution_providers/QNN-ExecutionProvider.md	Documents the new session config key.
onnxruntime/test/providers/qnn/matmul_test.cc	4 new tests: MatMulDisableFC_2D_StaticWeight, MatMulDisableFC_3D_StaticWeight, MatMulDefaultUsesFC, MatMulOptInUsesMatMul.
onnxruntime/test/providers/qnn/gemm_test.cc	3 new tests: GemmDisableFC_NoBias_TransB0, GemmDisableFC_NoBias_TransB1, GemmDisableFC_WithBias.