[QNN EP] Refactor LPBQ lowering helpers across op builders

[qti-ashimaj]

Description

Refactors the block-quantized (BQ) and LPBQ weight lowering logic across the QNN EP op builders (Conv, MatMul, MatMulNBits) to eliminate code duplication and consolidate shared helpers.

qnn_bq_utils :

• RegisterWeightAsConv1x1Filter: registers an LPBQ weight [K, N] as a [1, 1, K, N] HWCN 1×1 Conv2D filter, updating the LPBQ quant axis.
• AddConv2DNodeforBQLowering: creates a QNN Conv2D node (stride=1, pad=0, dilation=1, group=1) shared by both LPBQ and BwFloatBlock lowering paths.

qnn_utils :

• ProcessBiasForQuantizedOp: handles bias processing for quantized ops, requantizes INT32 bias if scales mismatch or quantizes float bias using activation_scale * weight_scale[c].

conv_op_builder : removed duplicated helpers (GetActivationScale, BuildBiasQuantParams, AddStaticBiasTensor, ProcessRequantizeBias, ProcessFloatBias).

matmul_op_builder / matmulnbits_op_builder : updated to use the new shared helpers.

matmul_test : enabled MatMulOp_QDQ_BlockQuant on Windows in addition to Linux.

Motivation and Context

The LPBQ lowering code was duplicated across various op builders like conv_op_builder, matmul_op_builder, and matmulnbits_op_builder. This refactor centralizes the shared logic into qnn_bq_utils and qnn_utils, making future changes (e.g., adding LPBQ support to new op builders) require only a single implementation.

[minfhong-qti]

Could you help rebase on mainline and resolve the conflicts?