feat(glm5-next): add GLM-5.3-Flash training support (#3699)

* feat(glm5-next): support GLM-5.3-Flash training

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* feat(glm5-next): add CP2 MedPix recipe

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm5-next): keep CP gather live for empty shards

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm5-next): keep validated EP72 CP2 recipe

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix: align GLM-5.3 checkpoint and MoE parity

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* chore(glm5-next): drop single-device loader change

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm5-next): keep example W&B opt-in

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* perf(glm5-next): use cuDNN sparse attention

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm5-next): guard optional imports and document model

Keep the base installation importable without VLM or FLA extras, add regression coverage for the torchvision-free path, and register the new architecture in model coverage docs.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm5-next): decay KDA state along key dimension

Match the pure-Torch recurrent KDA fallback to FLA's reference state update and cover the decay axis numerically.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(docs): place GLM-5.3 under publishing org

Follow the existing zai-org to thudm documentation slug mapping so recipe documentation coverage recognizes the model card.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

---------

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>