[GPU] Revise CUDA and ROCm plans around emulation-first evidence

[iteathen]

Summary

Revises the post-recovery GPU roadmap after current emulator/compatibility research.

• CUDA ([GPU][POST-RECOVERY][CUDA] Qualify CUDA emulation/compatibility first, then native NVIDIA execution #186): no-physical-GPU semantic emulation is now the preferred first feasibility step; GPGPU-Sim is the first candidate, followed by SCALE/ZLUDA CUDA-on-AMD compatibility where useful, then native NVIDIA for native-device/performance evidence.
• ROCm/AMD ([GPU][POST-RECOVERY][ROCm] Qualify rocJITsu emulation, VM device exposure, and native AMD execution #283): adds rocJITsu as the preferred first emulator candidate, explicitly separates its already-useful in-process path from the still-maturing vfio-user/QEMU virtual PCIe device path, keeps gem5 as a heavier fallback/reference, and retains physical AMD for native/performance proof.
• Generalized routing ([Feature][GPU] Generalize compute validity/routing across emulated, compatibility, and native backends #162): moves only the small neutral validity schema early; the broad detector/router remains later.
• Evidence now distinguishes compile, emulated-functional, compatibility/translated-device, native-device, hardware-specific, and performance validity.

Research basis

• ROCm rocJITsu vfio-user/QEMU work: [Feature]: rocjitsu: vfio-user GPU mode — expose rocjitsu as PCIe AMD GPUI via libvfio-user ROCm/rocm-systems#8218
• rocJITsu real ROCgdb + HIP-kernel emulator proof: feat(rocjitsu): drive GPU wave debugging end-to-end with ROCgdb ROCm/rocm-systems#8371
• GPGPU-Sim: https://github.com/gpgpu-sim/gpgpu-sim_distribution
• SCALE CUDA-on-AMD toolkit: https://docs.scale-lang.com/stable/
• ZLUDA: https://github.com/vosen/ZLUDA
• gem5 Full System AMD GPU model: https://www.gem5.org/documentation/general_docs/gpu_models/gpufs

Scope

Documentation only:

• docs/gpu-execution-profiles.md
• docs/roadmap.md

No runtime/provider/setup behavior changes in this PR.

Refs #162, #186, #283.