FIX: robust DatetimeIndex.union across DST transitions

[antznette1]

• Guard range-based setops for tz-aware DatetimeIndex to avoid RangeIndex assumptions that don’t hold across DST transitions in local time.
• For tz-aware union, compute union in UTC, then tz-convert back and infer freq. This preserves tz and correct ordering without fabricating nonexistent local times across DST.
• Flips new test to pass: test_union_across_dst_boundary.
• Notes:
  • Uses freq="h" alias.
  • Aligns expected unit via .as_unit(left.unit).

Ref: #62915 and tests-only PR #63087.

[jbrockmendel]

Before I spend more time on this, did a human write it?

[antznette1]

Before I spend more time on this, did a human write it?

Yes, this was written by me.
I see your point about _can_range_setop: by adding the tz restriction there I’m blocking the range fast path for a lot of cases, which is too broad. The DST bug I’m targeting only needs a special-case in the union logic for tz-aware DatetimeIndex, not a blanket change to _can_range_setop.