Extract render passes into RenderPass subclasses

[brendancol]

Follow-up to #73. The render graph framework (RenderPass, RenderGraph, CompiledGraph) landed in #76 but doesn't touch the actual render pipeline. This issue is about extracting the stages from render() and _update_frame() into concrete RenderPass subclasses.

Goal

Replace the hardcoded sequential pipeline in analysis/render.py and engine.py with render graph execution. Right now, adding or reordering passes means editing control flow by hand.

Passes to extract

Each pass wraps existing CUDA kernel calls. No new GPU code, just reorganizing.

Pass	Current location	Inputs	Outputs
GBufferPass	_generate_perspective_rays() + optix.trace() (primary)	scene, camera params	primary_rays, primary_hits, albedo, instance_ids, primitive_ids
ShadowPass	_generate_shadow_rays_from_hits() + optix.trace() (occlusion)	primary_rays, primary_hits, sun params	shadow_hits
AOPass	AO sample loop + _generate_ao_rays() + _accumulate_ao/gi()	primary_rays, primary_hits	ao_factor, gi_color
ReflectionPass	_generate_reflection_rays() + optix.trace()	primary_rays, primary_hits, instance_ids	reflection_hits
ShadePass	_shade_terrain()	all ray buffers, colormap, overlays	color, albedo
DenoisePass	denoise() from rtx.py	color, albedo, normals	denoised_color
EdgeOutlinePass	_edge_outline()	color, instance_ids	color (in-place)
EDLPass	_edl()	color, depth	color (in-place)
BloomPass	_bloom()	color	color (in-place)
TonemapPass	_tone_map_aces()	color	color (in-place)

Implementation approach

1. Start with render() (the offline path) since it's simpler — no accumulation loop, no particle splatting, no async readback.
2. Pass classes go in rtxpy/render_passes.py.
3. A build_default_graph() factory builds the same pipeline the current code runs.
4. Add use_render_graph=False to render() so the graph path is opt-in while both paths coexist.
5. Once the graph path matches existing output pixel-for-pixel, flip the default and remove the old code path.
6. Adapt _update_frame() in the viewer to the same graph, with particle splatting and accumulation as viewer-specific passes.

Watch out for

• In-place passes (EdgeOutline, EDL, Bloom, Tonemap) read and write the same buffer. The graph handles this (a pass can list a buffer in both inputs and outputs), but they have to run in the right order after shading.
• AO accumulation runs multiple samples per frame with progressive accumulation in _update_frame(). Probably best modeled as a single AOPass that loops internally, not N separate pass executions.
• External state — camera params, sun direction, colormap LUTs, overlay textures aren't graph buffers. Pass them via external_buffers or a config dict on the pass.
• Double allocation — _RenderBuffers currently handles allocation, the graph's AllocationPlan will replace it. During the opt-in period, can't have both allocating at once.

Done when

• render(use_render_graph=True) produces identical output to the current path (pixel-level comparison)
• explore() works with the graph path
• A custom pass can be inserted into the default graph (e.g. a user post-process)
• Disabling a pass via the graph (e.g. AO, denoise) matches disabling via render parameters