Research spike: indexed join strategy for v2 merge_insert

What to build

Time-boxed investigation to decide the approach for supporting scalar-indexed
joins in the v2 merge_insert path. Two approaches exist in the tracker:

• ScalarIndexJoinExec (#3480)
• Finger search (#4648)

Key questions

• Can the existing v1 indexed join logic be wrapped as a DataFusion exec node?
• Is a custom ScalarIndexJoinExec the right abstraction, or can we use
  DataFusion's built-in join with a custom table provider that reads from the
  index?
• Does finger search apply as an optimization within whatever node we choose?

See milestone description for full PRD.

Acceptance criteria

• Written design describing the chosen approach
• #3480 description updated with the chosen design (or a new issue created if the approach diverges)
• Key trade-offs documented (complexity, performance, reuse of existing code)

Blocked by

None - can start immediately.

User stories addressed

• 3: indexed upserts use v2 path (design phase)

Start by reading the milestone PRD, the existing v1 indexed join logic, and the designs in #3480 and #4648. Compare the possible DataFusion execution-node and table-provider approaches, including whether finger search fits, then document the chosen design, trade-offs, and update #3480 or create a replacement issue.