[Fix] Keep grouped token length estimates consistent

[primorLee]

Motivation

Fixes #1205.

When a conversation length was first encountered, the dataset appended the tokenizer-only length to self.length, but cached a larger value that also included the visual-token estimate. A later sample with the same string length therefore received a different estimate, making group_by_length order-dependent.

What changed

• add the existing visual-token estimate to token_length before both appending and caching it
• apply the same fix to the standard SFT, MPO, and pretraining loaders
• keep the GPT-OSS SFT and MPO copies aligned
• preserve the current approximate cache key and visual-token formula

Validation

• python -m py_compile on all five changed training entrypoints
• targeted pre-commit flake8 hook on all five files
• git diff --check