[ShilohYu]

Summary

Pure refactor (~40 lines of production code, plus 9 new unit tests). Three sites previously ran a full torch.cat copy even for single-element lists; they now reuse the same storage directly:

• glm53_vl_tokenize_fn — single-image / single-video samples reuse the processor output
• glm53_vl_sft_collator — single-media packs reuse the cast tensor
• Glm53ForConditionalGeneration._prepare_llm_inputs — single-modality packs reuse the SequenceContext block

Reuse does not change any value, dtype, or shape. Vision payloads are read-only end-to-end (SP split, H2D, patch_embed), and the "single-element reuse / multi-element still cat" behavior is pinned by 9 new unit tests.

On top of #2207 (early bf16), this PR removes the second copy of every single-media pack.

Measured results

• Microbenchmark (real training path): collate 9.24 → 4.62 ms (−50%), transient RSS delta 92.9 → 34.7 MB
• End-to-end (single node 8×GPU, 512K / SP8 / GBS1 / 25B, 10 steps): per-rank peak RSS 110.6 → 86.3 GiB (−22.0%)
• Numerical equivalence: step-1 loss bitwise identical across three runs; 10-step loss deviation below the same-code rerun noise floor; GPU memory bitwise unchanged (changes are CPU-side only)