Migrate GPT-OSS to HybridModel

[Phlip79]

What does this PR do ?

Migrates GPT-OSS from using GPTModel to HybridModel. To accomplish this, I needed to update HybridModelProvider to support yarn, which HybridModel already supports in MCore.

Perf comparison results

Needs #4508 for correct TFLOPS calculation

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Phlip79]

/ok to test c5e00e4

[Phlip79]

/ok to test 8e4d0fa

[claude]

Light Code Review

Overall the migration from GPTModel/GPTModelProvider to HybridModel/HybridModelProvider looks correct and well-tested for the core paths. A few observations:

1. Missing test for megatron_to_hf_config - The new classmethod that recovers num_hidden_layers from the hybrid layer pattern has no unit test. A small test with a known pattern would guard against regressions in the Symbols.PIPE/MTP_SEPARATOR stripping logic.

2. Dual model.norm.weight mapping - Intentional per the test, but deserves a brief inline comment explaining why both decoder.final_layernorm.weight and decoder.final_norm.weight are registered.

3. YARN settings not verified in test - test_provider_bridge_maps_config does not assert position_embedding_type == yarn or any YARN-specific field, even though extending HybridModelProvider with YARN fields is a key part of this PR.

---

Suggested test cases: No perf/recipe config files are changed in this PR, but this migration affects all GPT-OSS perf tests since the underlying provider type changed. All GPT-OSS perf configs (20B and 120B, all GPU targets and precisions) should be validated: gpt_oss_20b_8gpu_pretrain_perf (all GPU/precision combos), gpt_oss_120b_pretrain_perf (all GPU/precision combos), test_gpt_oss_120b_perf_config_instantiation, L1 functional L1_Launch_recipes_gpt_oss (pretrain + finetune), L1 functional L1_Launch_models_gpt_oss (model conversion tests).