Unify accuracy-mode configuration and preserve explicit clipping

[zrr1999]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Use the existing use_accuracy_compatible argument for model and optimizer numerics. Delete the SFT initialization override that silently sets clip_grad to zero, and remove the separate reproducible-norm and native-unfused-AdamW arguments. DSA padding and bridge behavior follow the model architecture and shared mode. An older Megatron optimizer configuration that cannot honor this mode fails explicitly.

Requires Megatron #17; PaddleFleet #1961 removes the retired options from the GLM cases. Reference wheels must include both companion changes.

Experiment results

• 27 focused tests pass, including all requested clipping thresholds in both modes, configuration propagation, ordinary-model padding, DSA bridge behavior, and collator construction before model_info.config is populated.
• Final owning heads ran native Paddle and Torch entrypoints for 100 steps in formal-ci-followup-20260915-n100-r1: all 100 loss values are IEEE bit-identical, provenance passes, and both canonical safetensors checkpoints match. Each side also matches its prior accepted 100-step loss and checkpoint. GLM4.5 native EP2 10-step regression has 40/40 matching loss hashes and matches the previous baseline on both sides. Local receipts: experiments/ops/ci-followup-20260915/terminal.json and glm45-pair-r0/comparison.json. Frozen repository pins, existing completion-required items and the incomplete native-build snapshot remain unresolved; this is candidate acceptance, not complete + sealed or remote-CI acceptance.
• Flake8, isort and YAPF pass on the changed files; the final padding change also passes the repository pre-commit hooks under CI Python 3.10.