NHWC 1x1 conv with padding triggers `CpuGemmLowpMatrixMultiplyCore::validate` mismatch

[alvoron]

When using NHWC, 1x1 kernel, stride=1, and non-zero padding, NEConvolutionLayer::validate can fail with:

  "in validate src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp:524: a->dimension(1) != output->dimension(1)"

This happens because the GEMM3D fast path (skip_im2col/skip_col2im) assumes output spatial equals input spatial.
With padding, output W/H increase, violating that assumption.

Reproducer below triggers NEConvolutionLayer::validate to return this error:

#include "arm_compute/core/Error.h"
#include "arm_compute/core/TensorShape.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/runtime/Tensor.h"
#include "arm_compute/runtime/NEON/NEFunctions.h"
#include <iostream>

using namespace arm_compute;

int main() {
    // Reproducer for CpuGemmLowpMatrixMultiplyCore::validate failure when
    // skip_im2col/skip_col2im is used with NHWC 1x1 conv + padding.
    TensorInfo src_info(TensorShape(1, 100, 128, 1), 1, DataType::QASYMM8, DataLayout::NHWC);
    TensorInfo wei_info(TensorShape(1, 1, 1, 64), 1, DataType::QASYMM8, DataLayout::NHWC);
    TensorInfo dst_info(TensorShape(64, 102, 130, 1), 1, DataType::QASYMM8, DataLayout::NHWC);

    src_info.set_quantization_info(QuantizationInfo(1.0f, 0));
    wei_info.set_quantization_info(QuantizationInfo(1.0f, 0));
    dst_info.set_quantization_info(QuantizationInfo(1.0f, 0));

    PadStrideInfo conv_info(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR);
    ActivationLayerInfo act_info;
    WeightsInfo weights_info;

    auto status = NEConvolutionLayer::validate(&src_info, &wei_info, nullptr, &dst_info, conv_info, weights_info, Size2D(1, 1), act_info);
    if (!status) {
        std::cout << "VALIDATE FAILED: " << status.error_description() << std::endl;
        return 1;
    }
    return 0;
}

As a potential fix, I've added padding and dimension check in skip_im2col initialization in CpuGemmConv2d::skip_im_col_info():

    const bool has_padding   = conv_info.pad_left() != 0 || conv_info.pad_right() != 0 ||
                               conv_info.pad_top() != 0 || conv_info.pad_bottom() != 0;
    const bool same_spatial  = (conv_w == src->dimension(idx_width)) && (conv_h == src->dimension(idx_height));
    const bool skip_im2col   = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 &&
                               conv_info.stride().first == 1 && conv_info.stride().second == 1 &&
                               !has_padding && same_spatial);