rendering: reuse retained CPU work

[tritao]

Builds on #78.

Purpose

After unchanged-frame reuse and incremental updates remove the large rebuilds,
the remaining CPU cost is dominated by smaller repeated work: constructing and
copying command data, allocating short-lived containers, resolving the same
resource/cache indices, and rebinding state that has not changed.

This PR removes those costs from the existing retained pipeline without
changing its semantic model or adding new public APIs.

What changed

• Move completed retained commands directly into draw-list storage.
• Streamline command construction and compact retained command paths.
• Use internal inline storage for small hot-path collections.
• Reuse command-to-resource and prepared cache indices during submission.
• Preserve compatible visual submission, shader, material-uniform, depth, and
  blend state across adjacent commands.

Scope

These are internal CPU hot-path optimizations. The inline-storage helper is
not installed or exposed as a public Coin container, and public retained-IR
value types keep their standard-container API. Planner ordering and mutation
semantics remain owned by the lower layers.

Validation

• Retained node and material/lighting correctness tests.
• Existing visual, picking, and selection coverage across the optimized
  submission path.
• Benchmark phase data verifies reduced construction and command-execution
  work without changing semantic draw counts or output.