STYLE: STAPLEImageFilter counts pixels in a double accumulator instead of GetNumberOfPixels()

[hjmjohnson]

STAPLEImageFilter::GenerateData() counts pixels in a double accumulator rather than reading the region's pixel count directly:

https://github.com/InsightSoftwareConsortium/ITK/blob/main/Modules/Filtering/ImageCompare/include/itkSTAPLEImageFilter.hxx#L110-L125

double N = 0.0;
double g_t = 0.0;
{
  ImageScanlineIterator out(W, W->GetRequestedRegion());
  while (!out.IsAtEnd())
  {
    while (!out.IsAtEndOfLine())
    {
      out.Set(out.Get() / static_cast<double>(number_of_input_files));
      g_t += out.Get();
      N = N + 1.0;          // <-- pixel counter held in double
      ++out;
    }
    out.NextLine();
  }
  g_t = (g_t / N) * m_ConfidenceWeight;
}

This is not a live defect. double has a 53-bit mantissa, so the increment stays exact up to 2^53 pixels — unreachable. Filing it as cleanup only, because it is the same idiom as the float counter fixed in #6759, and someone auditing for that pattern will land here and have to re-derive that it's benign.

Suggested change: drop N and use the region's pixel count, which is exact by construction and removes a per-pixel add from the inner scanline loop.

const SizeValueType N = W->GetRequestedRegion().GetNumberOfPixels();

g_t itself must stay floating-point — it accumulates fractional per-pixel values, not counts.

Scope check — what else in this file was examined

• The initial averaging loop out.Set(out.Get() + 1.0) accumulates in the output pixel type, but is bounded by number_of_input_files, so it carries no precision risk.
• p_num / p_denom / q_num / q_denom accumulate fractional weights W_i and 1 - W_i over all pixels. These are genuinely fractional and already double; they are not counters and should not be converted.
• alpha1 / beta1 are per-pixel products, not accumulations.

itkLabelVotingImageFilter was also checked and already counts votes in unsigned int.

Related: #6759 (the float variant of this pattern in MultiLabelSTAPLEImageFilter, where it was a real defect — float saturates at 2^24 = 16,777,216, which a 3D image reaches easily).