v2: Swap shadow_raycaster for NREL bifacial_radiance

[mokahlo]

Goal

Replace the simplified 2D shadow_raycaster in models/shadow_raycaster.py with NREL's bifacial_radiance for research-grade accuracy.

Why

Our current raycaster does a 2D projection of the panel polygon onto the ground and intersects with the bed rectangle. This is good enough for v1 but misses:

• Bifacial back-face gain — we approximate as a flat 10%. Real value depends on ground albedo (bed soil) and rear-side illumination. bifacial_radiance handles this properly.
• Ground-reflected light to bed — bed DLI currently treats panel shadow as total block. Real DLI includes diffuse + reflected light even in shade.
• Time-of-day geometry — our projection assumes thin panel. bifacial_radiance does proper 3D scene ray-tracing.

Reference

• NREL bifacial_radiance: https://github.com/NREL/bifacial_radiance
• NREL InSPIRE tutorials: https://github.com/NREL/InSPIRE

Scope

• Install bifacial_radiance (Radiance engine required, non-trivial on Windows)
• Define Wattplot scene: bed rectangle, panel tilted at 0-90 deg, soil as ground with measured albedo
• Run baseline simulation for Phoenix 2025 with 5 tilt schedules
• Compare DLI and kWh from bifacial_radiance vs our shadow_raycaster
• If significantly different, swap the simulator to use bifacial_radiance
• Document the new sim approach in README

Done when

Differences between our raycaster and bifacial_radiance are documented and the more accurate one is in use.