feat(bindings): spanUnion(span | spanset) → spanset aggregate

[estebanzimanyi]

Summary

Adds `spanUnion` aggregate — merges a column of spans (or spansets) into a single canonical spanset. 10 overloads:

Input	Output
`spanUnion()`	`` for `{int, bigint, float, date, tstz}span` (5)
`spanUnion()`	`` same 5 types

Closes the architectural gap documented in the parity manifest at `test/sql/parity/015_span_aggfuncs.test` (PR #21).

```sql
SELECT spanUnion(s) FROM (VALUES (intspan '[1, 5)'), (intspan '[3, 8)')) tt(s);
-- {[1, 8)} -- overlap merged

SELECT spanUnion(s) FROM (VALUES (intspan '[1, 3)'), (intspan '[3, 5)')) tt(s);
-- {[1, 5)} -- adjacent merged

SELECT spanUnion(s) FROM (VALUES (intspan '[1, 5)'), (intspan '[10, 12)')) tt(s);
-- {[1, 5), [10, 12)} -- disjoint preserved
```

Implementation

New `SpansetUnionState` holds a heap-allocated `SpanSet *`. Same destructor pattern as the skiplist-backed aggregates (PR #50, PR #52).

Step	What it does
`Operation`	Copies input blob, calls MEOS `span_union_transfn` (or `spanset_union_transfn` when `INPUT_IS_SPANSET=true`), reassigns `state.spanset` to whatever the transfn returns — the transfn may return a new larger spanset and free the old state when it grows.
`Combine`	Calls `spanset_union_transfn(target, source)` to merge two states.
`Finalize`	Calls `spanset_union_finalfn` which compacts and frees the state. We null `state.spanset` after to keep the destructor from double-freeing.

The transfn dispatch is templated on `bool INPUT_IS_SPANSET` so the same `SpanUnionFunction` handles both Span and SpanSet inputs.

Stacked on

• feat(bindings): aggregate function infrastructure — extent(span) for all 5 span types #47 → feat(bindings): extend extent(*) to spanset and set surfaces #48 → feat(bindings): extent(tnumber) → tbox + bundle extent overloads in one set #49 → feat(bindings): tCount(temporal) aggregate — first skiplist-backed aggregate #50 → feat(bindings): tAnd / tOr / tagg_min / tagg_max / tagg_sum aggregates #52 (aggregate stack)

Test plan

• Distinct spans → preserved as multiple elements
• Overlapping spans → merged
• Adjacent spans → merged
• Spanset inputs (multiple element merges)
• All 5 base types (int / bigint / float / date / tstz)
• NULL handling: mixed NULL + value → returns the spanset; all-NULL → returns NULL
• No segfaults (destructor + finalfn double-free guard)

[estebanzimanyi]

Consolidated into #60 (aggregate cluster squash). The full chain of aggregate work is now reviewable as a single PR; this branch's commits are preserved in #60's history. Closing to minimize the review queue.