Support Helion RoPE override for ComplexRoPE (#3767)

This PR adds a HelionComplexRoPE override targeting exact
ComplexRoPE.Config, enabling DeepSeek V3 through the existing
`--override.imports torchtitan.overrides.helion_rope` path.

While Helion doesn't support complex tensors, it instead produces the
cos and sin output separately and puts them together as a complex tensor
in torch. Similar to cos_sin rope, we also add an integration test for
deepseek v3.

We benchmark on 2xGB200:
- DeepSeek V3 debugmodel, fwd_bwd only
- DP shard=2, EP=2, TP=1, CP=1, PP=1, full_dtensor
- local batch size 4, global batch size 8
- warmup=5, iters=20
- model compile off
  - FlexAttention compile on
  - fwd_bwd (in ms)
  - we compared loss and gradient l2 norms to confirm accuracy.

| seq_len | baseline | with Helion ComplexRoPE | speedup |
|---:|---:|---:|---:|
| 2048 | 563.246 ms | 511.982 ms | 1.100x |
| 4096 | 570.322 ms | 526.500 ms | 1.083x |
| 8192 | 572.380 ms | 542.133 ms | 1.056x |
| 16384 | 568.548 ms | 539.765 ms | 1.053x |
| 32768 | 636.378 ms | 625.146 ms | 1.018x |
| 65536 |1107.151 ms | 1090.194 ms | 1.016x |

Co-authored-by: Ethan Che <eche@meta.com>

Author: ethche

tests/integration_tests/models.py

@@ -59,11 +59,15 @@ def build_model_tests_list() -> list[OverrideDefinitions]:
                     "--parallelism.data_parallel_shard_degree 4",
                     "--parallelism.expert_parallel_degree 2",
                     "--compile.enable",
+                    "--override.imports torchtitan.overrides.helion_rope",
                 ],
             ],
-            "DeepSeek V3 FSDP+EP+compile",
+            "DeepSeek V3 FSDP+EP+compile (+ Helion RoPE override)",
             "deepseek_v3_fsdp+ep+compile",
             ngpu=4,
+            # The Helion fused RoPE kernels are CUDA-only and tuned for NVIDIA
+            # H100/GB200; skip on ROCm where they are unvalidated.
+            skip_rocm_test=True,
         ),
         OverrideDefinitions(
             [