[Fix] Correct generic MTP future embeddings and empty-shard loss

[jayhenry]

Stack #2179 (bottom to top):

1. [Fix] Preserve DeepEP layout events in synchronous dispatch #2177 fix/deepep-layout-event → main
2. [Feature] Add GLM-5.3-Flash F0: 25B cropped reference checkpoint builder #2105 feat/glm53flash-materialize-full-f0 → fix/deepep-layout-event
3. [Feature] Add GLM-5.3-Flash F3: Kimi Delta Attention (KDA) #2106 feat/glm53flash-f3-kda → feat/glm53flash-materialize-full-f0
4. [Feature] Add GLM-5.3-Flash F4: mHC four-stream residual #2107 feat/glm53flash-f4-mhc → feat/glm53flash-f3-kda
5. [Feature] Add GLM-5.3-Flash F5: NoPE DSA + KPool indexer + clamped SwiGLU #2108 feat/glm53flash-f5-nope-dsa → feat/glm53flash-f4-mhc
6. [Feature] Add GLM-5.3-Flash F1: VL data preprocessing pipeline #2109 feat/glm53flash-f1-vl-data → feat/glm53flash-f5-nope-dsa
7. [Feature] Add GLM-5.3-Flash F2: vision tower + projector (eager) #2110 feat/glm53flash-f2-vision-tower → feat/glm53flash-f1-vl-data
8. [Fix] Correct generic MTP future embeddings and empty-shard loss #2178 fix/mtp-local-future-embeddings → feat/glm53flash-f2-vision-tower ← you are here
9. [Feature] Add GLM-5.3-Flash F6 core: text model + MTP + compose model #2111 feat/glm53flash-f6-text-moe → fix/mtp-local-future-embeddings

Base is #2110's branch (layer 7). Review only this PR's own diff.

---

Summary

• Gather input embeddings once and build only the local packed future slice at each MTP depth, preserving the existing gradient path and packed sequence boundaries.
• Keep the empty SP shard loss in FP32 while retaining zero gradient edges for MTP outputs.
• Add public behavior tests for values, gradients, checkpoint replay and empty shards.

This generic MTP layer follows the F2 vision layer. GLM-5.3-Flash specific MTP wiring and integration tests remain in F6.

Tests

• tests/loss/test_mtp_e2e_tv_loss.py, tests/loss/test_mtp_empty_shards.py and tests/model/test_mtp_local_future_embeddings.py passed in the targeted CPU run.
• The GLM53 MTP integration case passed in the separate 17-test GPU run.