The previous range-based for loop called connections.erase(it.first) on
the current element and then continued iterating - undefined behavior,
because erase invalidates the iterator the range-based for is holding.
Switch to a manual iterator loop that uses erase's return value to
advance past the removed element. Update connection_count once at the
end instead of on every erase.

The INT64 case in ColumnArrowToDuckDBRunEndEncoded incorrectly
dispatched to FlattenRunEndsSwitch<int32_t> instead of
FlattenRunEndsSwitch<int64_t>, a copy-paste error from the INT32 case.
This caused corrupted data when scanning Arrow REE columns with int64
run_ends, as the int64 buffer was reinterpreted with int32 stride.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…orruption

Apply GetEffectiveOffset() to the union type_ids buffer read, matching
the pattern used by every other Arrow type in ColumnArrowToDuckDB.

Without this fix, when scanning Arrow chunks larger than
STANDARD_VECTOR_SIZE (2048 rows), subsequent scan passes re-read
type_ids from offset 0 instead of advancing by chunk_offset, pairing
wrong type tags with correct child data — silent data corruption.

Also fixes Arrow union appender OOB write when `from > 0` and adds
comprehensive UNION type roundtrip tests to arrow_roundtrip.cpp.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

See duckdb#22348

See duckdb#21983 for original PR. This
also required to backport duckdb#20333
first.