[AMD] ci: enable the GPT-OSS 20B MoE LoRA e2e on ROCm

[XinyuJiangCMU]

Co-author-with: @JessicaJiang-123

Summary

Register tests/e2e/megatron/model_scripts/test_gpt_oss_20b_moe_lora_ci.py on the ROCm stage-c-4-gpu-mi350 suite. The test is otherwise unchanged; both serving combinations already run as written.

The one behaviour change is that ROCm selects SGLang's triton attention backend, gated on MILES_HARDWARE_PLATFORM. ROCm otherwise defaults to aiter, which ships no compiled batch-prefill kernel for what gpt-oss-20b needs: a learned attention sink, a KV cache large enough to require the global-load variant, and page_size=1 below the kernel's tile width. That variable is only set in docker/Dockerfile.rocm, so the argument string is empty on every CUDA image and the composed args are byte-identical there. The CUDA registration is untouched.

Verification

Ran the test end to end on 4x MI355X with both dependencies below applied. Each combination completed and the Ray job reported success:

combo PASSED: shared-outer + virtual-experts
combo PASSED: per-expert + no-virtual-experts
update_weights phase=end ok=true elapsed_s=4.6
Job 'raysubmit_wmNUyY7nFiEgEPcu' succeeded

collect_tests(discover_ci_files(), sanity_check=True) returns both a HWBackend.CUDA entry on stage-c-4-gpu-h200 and a HWBackend.ROCM entry on stage-c-4-gpu-mi350 for this file, and pre-commit run passes.

Dependencies

Merge this last; ahead of either one it turns the mi350 suite red.

• [AMD] support ROCm TE grouped expert LoRA Megatron-Bridge#29 makes the grouped-expert LoRA adapter work against ROCm's TransformerEngine. There is no PR-description directive for Bridge: docker/Dockerfile.rocm installs it from the bridge branch head, so the ROCm image needs a rebuild after it merges.
• [AMD] Fix the MoE LoRA JIT compilation on ROCm sgl-project/sglang#34834 makes the MoE LoRA align JIT kernel build on ROCm, which the per-expert combination needs.