Fix batch training padding mismatch handling

[H-Chris233]

Motivation

• Batch training used padded Batch inputs while create_training_batches produced targets based on real token lengths, causing logits/target length mismatches and incorrect gradient application in train_monitored_batch.
• Improve clarity for this educational project by documenting the per-sample real-length slicing approach so learners understand why PAD tokens must not participate in training.

Description

• In src/llm.rs train_monitored_batch compute per-sample real_len from input_batch.attention_mask and slice input tokens to tokens[0..real_len] so forward/backward operate only on real tokens, collecting per-sample batch_targets accordingly.
• Add a safety check to skip malformed samples when log_probs.nrows() != target_ids.len() and avoid crashing or applying mismatched gradients.
• Remove the previous approach of zeroing gradients by PAD positions because samples are now sliced to exclude PAD before gradient computation.
• Update docs/批量训练与动态掩码.md to explain the real-length slicing rule and why PAD must not participate in gradient computation.

Testing

• No automated tests were executed as part of this change.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3afa7178c7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".