Add model-aware support for MI300X, MI325X, and MI300A

[irvineoy]

Summary

• compose shared CDNA 3 / gfx942 guidance with separate MI300X, MI325X, and MI300A hardware profiles
• recognize all three exact model names in runtime image and ROCm architecture resolution while retaining the generic MI300 fallback
• add model-specific quickstarts and pass the same architecture context to mini_swe_triton
• document model-specific limits and distinguish configuration coverage from pending physical-GPU validation

Tests

• make check-gpu-architecture
• make check-docker-runner
• make check-evaluator
• make check-held-out
• make check-visualization
• make check-perf-helpers
• Python compile checks and shell syntax checks