build(cuda)!: raise minimum CUDA support to 12.6

[plexoos]

Raises the minimum supported CUDA version from 12.1 to 12.6 and updates CI/container coverage to
avoid CUDA compiler versions that hang while compiling optimized Release PTX.

PR #380 exposed a CUDA compiler issue when CSGOptiX7.cu is compiled as optimized Release PTX with production definitions. Older CUDA 12.x compilers enter a high-CPU cicc hang while trying to optimize/inline the generated device code.

Local testing showed:

CUDA	Release PTX result
12.5.1	hangs / timeout
12.6.3	hangs / timeout
12.8.2	passes
12.9.2	passes
13.0.2	passes in CI

CUDA 12.6 is kept as the general minimum supported version, while Release image coverage uses CUDA 12.8.2+ because CUDA 12.6.3 still reproduces the Release PTX compiler hang.

• Require CUDA 12.6+ at top-level CMake configure time.
• Require CUDAToolkit 12.6 in package config and CUDA-dependent test CMake files.
• Replace legacy CUDA 12.1/12.5 CI images with:
  • CUDA 12.6.3 for minimum develop/debug coverage.
  • CUDA 12.8.2 for oldest Release coverage.
• Update README and getting-started docs to document CUDA 12.6+.
• Document CUDA 12.8.2+ as recommended for production Release PTX builds.

CUDA 12.1 and 12.5 are no longer supported by the project CI matrix or CMake configuration.

[plexoos]

There is no immediate need to raise the minimum CUDA version to 12.6. The issue discussed in #380 has been addressed there by forcing qsim::propagate(...) to be inlined.