[Elvand-Lie]

This Pull request:

Changes or fixes:

TProfile2D::Divide and TProfile3D::Divide return wrong statistics for the Y and Z axes:

TProfile3D num("num", "", 4, 0, 4, 5, 0, 10, 3, 0, 30);
TProfile3D den("den", "", 4, 0, 4, 5, 0, 10, 3, 0, 30);
TProfile3D res("res", "", 4, 0, 4, 5, 0, 10, 3, 0, 30);
num.Fill(0.5, 1., 25., 2.);
den.Fill(0.5, 1., 25., 1.);
res.Divide(&num, &den);
res.GetMean(3); // 5 on master; the only filled bin is at z = 25

If res already held a point at (3.5, 9, 5) before the division, master gives GetMean(2) = 5.5 and GetMean(3) = 7.5 instead of 1 and 25. TProfile2D gives GetMean(2) = 5.5 in the same situation.

There are three causes:

• Both Divide overloads of TProfile2D and TProfile3D reset only fEntries, fTsumw, fTsumw2, fTsumwx and fTsumwx2. The loop then adds the new Y and Z contributions (in 3D also XZ, YZ and T) to the old sums. Divide(h1) always divides a filled profile, so it is always affected. The 1D TProfile::Divide(h1) already resets all of its sums.
• TProfile3D::Divide(h1, h2) takes the Z bin center at the Y bin index: fZaxis.GetBinCenter(biny).
• TProfile3D::Divide(h1) accumulates fTsumwz += u and fTsumwz2 += u*z instead of u*z and u*z*z.

The fix resets every sum each loop accumulates, uses binz, and corrects the two Z moments.

Four tests in hist/hist/test/test_TProfile.cxx cover both classes and both overloads. They divide into filled targets, use axes whose X, Y and Z bin centers differ, and check the Z standard deviation.

• The four new tests fail on master and pass with this change.
• ctest -R '^gtest-hist-hist-' passes: 32 tests, 0 failures, on Linux with a minimal build.
• Not run: the full ROOT test suite, roottest, and platforms other than Linux.

Checklist:

• tested changes locally
• updated the docs (if necessary)

Codex (gpt-6.1-sol) implemented the fix and the tests; Claude Code (claude-opus-5-5) found the bugs and validated them.