rendering: accelerate retained interaction

[tritao]

Builds on #73.

Purpose

Hover picking, selection changes, and depth-stack queries can run far more
often than the scene itself changes. Rebuilding the same interaction work or
waiting synchronously for avoidable GPU readback makes those operations scale
with the full retained frame instead of the requested interaction.

This PR prepares and reuses interaction-specific work, adds a direct
frontmost-target path for closest picks, and introduces internal asynchronous
readback machinery while preserving the existing synchronous public API.

What changed

• Use the frontmost retained target for closest-pick queries.
• Prepare and reuse interaction plans.
• Batch selection and depth-stack submissions where ordering permits.
• Dispatch OpenGL synchronization objects through Coin's glue layer.
• Support asynchronous request, poll, and completion internally.
• Add interaction phase diagnostics and scaling workloads.

API and fallback

SoRenderManager remains synchronous. Backend request tokens, synchronization,
and prepared interaction caches remain implementation details until an actual
consumer establishes the requirements for a public asynchronous API.

Unsupported or stale prepared work falls back to the established interaction
rendering and readback paths.

Validation

• Closest, region, and overlapping depth-stack picking.
• Selection updates and batched interaction submissions.
• Synchronous and internal asynchronous result parity.
• Interaction latency and scaling curves across compatibility and core
  profiles.