[Tracking] E2E training alignment and convergence against Megatron/Megatron-Bridge

We should add an end-to-end training alignment test suite that compares AutoModel against Megatron/Megatron-Bridge under identical settings. The goal is to catch subtle correctness issues in the full training path, including tokenizer/chat rendering, sample packing, answer-only loss masks, reasoning-content masks, position ids, forward numerics, backward numerics, and optimizer-step behavior.

Recent answer-only loss-mask issues in multiturn chat rendering show that component-level tests are not enough. Some bugs may not reliably surface as obvious training failures, but can still change the effective objective or training dynamics. We need a reference-based alignment workflow for representative model families.

Scope

Compare AutoModel training against Megatron/Megatron-Bridge for these model families:

• Qwen3
• Qwen3.5
• Gemma4
• DeepSeek v4
• GLM 5.2

The test should cover the full training process, not only chat template rendering.