Add render graph framework

[brendancol]

Closes #73

Summary

Render graph framework for #73 (Phase 2 roadmap). Lets you define render pipelines as a DAG of passes instead of managing buffer dependencies and execution order by hand in render().

What's included

• RenderPass base class — typed buffer inputs/outputs per pass, optional capability requirements
• RenderGraph container — pass registration, fallback wiring, compile-to-execute
• Topological sort (Kahn's algorithm, stable tie-breaking by insertion order, cycle detection)
• Buffer lifetime analysis — greedy interval colouring assigns pool slots, reuses GPU memory when lifetimes don't overlap
• Capability gating — passes declare requirements (e.g. optix_denoiser); graph skips unavailable ones and wires fallbacks
• Validation — missing inputs, cycles, fallback chain resolution

Not included (follow-up)

Concrete pass extractions (GBuffer, Shadow, AO, Denoise, Tonemap, Composite wrapping existing CUDA kernels) are deferred. This PR is the framework they'll plug into. render() and _update_frame() are unchanged.

Test plan

• 38 unit tests across all framework components (topology, cycles, capability gating, fallbacks, lifetime analysis, execution, edge cases, realistic pipeline shape)
• Example script runs on CPU with numpy, no GPU needed
• Manual check that existing tests still pass (no changes to existing code paths)