BUG: ewm(adjust=False) NaN gaps disagree for times vs no-times

[hasancankeles]

• closes BUG/API: ewm with equally spaced times disagrees with no-times for adjust=False, ignore_na=False on NaN data #66523 (blocked on Needs Discussion / /take; xref until the label is removed)
• also addresses ewm NaN handling is wrong : simple example #31178 (same root cause: new_wt = 1 - old_wt used accumulated decay)
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.
• I have reviewed and followed all the contribution guidelines
• If I used AI to develop this pull request, I prompted it to follow AGENTS.md.

The adjust=False loop special-cased com == 1 (GH-59142) as new_wt = 1 - old_wt. That is the right convex weight when old_wt is just this step's decay, which is true for irregular times with no missing values. After an ignore_na=False NaN row, old_wt has already been multiplied once, so the same assignment gives the new observation the entire forgotten mass. That is:

• a 3.0 regression on the no-times path for com=1 (ewm NaN handling is wrong : simple example #31178: 4.0 instead of 11/3);
• the same wrong value on the times path, which always forces _com = 1.0.

Gating the special case on use_deltas (GH-66521) would fix only the first and split the two spellings — the inconsistency filed as #66523.

The intended split, from the documented ignore_na=False recursion, GH-59142's irregular-interval tests, and test_ewma_with_times_equal_spacing, is:

• weight the new observation by this step's interval: new_wt = 1 - (1-α)^δ (δ = 1 without times);
• leave decay accumulated across NaN rows in old_wt and renormalize.

That is option 1 from the issue, without switching the irregular no-NaN times path to the normalized form. test_ewma_with_times_variable_spacing is unchanged (0.2376251864…). Equally spaced times again match no-times for every adjust × ignore_na combination, including on NaN data.

Verified: pandas/tests/window/test_ewm.py, test_numba.py, test_online.py, and the other ewm-related window tests.

[github-actions]

Thanks for the pull request, @hasancankeles! It's linked to #66523, but that issue isn't assigned to you yet. To make sure two people don't unknowingly work on the same thing, we ask contributors to claim an issue first. Just comment /take on #66523 to claim it, and you're good to go. See the contributing guide for the full flow.