E2e challenge july fixes

[mwatson-nvidia]

• Improves NuPlan trajectory precision: translates large world/UTM coordinates in float64 before storing local float32 trajectories, avoiding sub-meter precision loss for ego and traffic agents.
• Updates MTGS scene/asset handling:
  • supports explicit scene-ID → asset-folder mappings;
  • makes the scene-data-source cache a bounded LRU cache controlled by --cache-size;
  • accepts legacy MTGS assets without road_height_map, logging a warning instead of failing.
  • adds coverage for mapping, cache eviction, and legacy assets.
• Adjusts NuPlan challenge runtime defaults:
  • builds gsplat CUDA kernels for sm_90 plus PTX;
  • enables aggregated rendering;
  • switches route generation from recorded to map-based, with a 40 m start offset / up to ~80 m lookahead;
  • restores four runtime workers for EC2 competition runs.
• Fixes low-speed linear MPC behavior: removes incorrect steering/velocity product derivatives that caused a discontinuity at the kinematic-model speed threshold. Adds regression tests for
  threshold continuity and steady turning.
• Updates renderer defaults to NRE 26.04, uses pycena_nrm_full, and enables the harmonizer.
• Adds a GitHub issues link to the E2E challenge README.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.