feat: add GLM-5.3-Flash MTP support

[cubezhang]

feat: add GLM-5.3-Flash MTP support

GLM-5.3-Flash checkpoint layer 45 was previously discarded, so the model could
only serve without speculative decoding. This change registers a GLM-specific
MTP model, reuses the existing NextN predictor prologue, and constructs layer 45
with GLM's k-pool MLA and clamped SwiGLU MoE behavior.

The pooled indexer now preserves position-addressed key/gate history across
speculative verification and rejection, including ragged concurrent requests.
Draft graph staging handles GLM-5.3's two-dimensional NextN residual even though
the backbone config contains hc_mult=4. The existing GLM-5.2/DeepSeek MTP path
keeps its original decoder behavior through the default predictor-layer hook.

Validation

• Black formatting and Ruff checks pass for the changed MTP implementation and tests.
• 61 focused GLM/MTP tests pass, including the shared prologue, input masking,
  top-k sharing, decode contracts, scheduling, GLM routing, and k-pool behavior.
• The final physical KV-slot mapping kernel test passes on ROCm.
• A standalone image built directly from upstream/main plus this commit
  completed real-model TP8 MTP3 text serving and the 8/8 quality smoke suite on
  8× MI308X.
• MTP1/MTP3 natural acceptance lengths 0–3, forced rejection, and concurrent
  short/long requests were validated during bring-up.

Measured performance

On a fixed natural-language workload with TP8, 512 output tokens, and BF16
KV/index cache:

Configuration	Concurrency	Output tok/s	Change vs no MTP
No MTP	1	104.67	—
MTP3	1	181.54	+73.44%
No MTP	8	409.96	—
MTP1	8	559.98	+36.59%

This branch is based directly on upstream/main (5a9c206).

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2157 --add-label <label>

[yitingw1]

Please fix conflicts

[cubezhang]

Okay,let me check.