[AMD] Enable amd pr ci

[JessicaJiang-123]

Co-author-with: @XinyuJiangCMU, @yuankaichen-amd, @Xinyu-Kang, @indianspeedster, @sreerohi, @lizamd

Summary

Bring up the ROCm CI lane on MI300X/MI355X and enable the cases, without changing shared NVIDIA behavior. ROCm-specific recipes live under scripts/amd/. Tests select them through MILES_PLATFORM=rocm, while smaller ROCm-only differences are passed through backend-neutral extra_args / extra_env_vars.

Test status

case	status	blocked on
test_amd_mtp1_spec_v2_r3	disabled	—
test_inkling_small_4layer_ci	disabled	—
test_qwen3_4B_offload_disk_stream	disabled	—
test_deepseek_v4_flash_4layer_ci	disabled	sgl-project/sglang#34330
test_qwen3_4B_fsdp_hybrid_shard_r2s2	disabled	#2265
test_amd_moriep_fp8_bridge	disabled	sgl-project/sglang#34006, sgl-project/sglang#34203, sgl-project/sglang#34220
test_glm5_2_744b_a40b_5layer_ci	disabled	sgl-project/sglang#34223
test_glm47_flash/test_amd_r3_mtp	disabled	ROCm/aiter#4653, sgl-project/sglang#34223, sgl-project/sglang#34238

How AMD coverage is added

ROCm-specific behavior is selected through MILES_PLATFORM=rocm, set by docker/Dockerfile.rocm. Shared NVIDIA paths do not contain ROCm-specific branches.

There are three patterns:

1. Different recipe → use scripts/amd/

If ROCm needs a different recipe, the test dispatches to an AMD-specific script:

if os.getenv("MILES_PLATFORM") == "rocm":
    from scripts.amd.run_inkling import _MODEL_REGISTRY, ScriptArgs, _train
else:
    from scripts.run_inkling import _MODEL_REGISTRY, ScriptArgs, _train

2. Same recipe, extra flags → use CaseConfig

If ROCm only needs extra arguments or environment variables, pass them through backend-neutral fields:

extra_args: str = ""
extra_env_vars: dict[str, str] = field(default_factory=dict)

3. AMD-only case → keep it in the AMD test

If the case only exists on ROCm, keep its configuration entirely in the AMD test.

[guapisolo]

This shouldn't be modified like this. Tom has a follow-up PR fix this.