[GLM5.2] Use official IndexShare support

[yapdianang]

Goal

Train GLM-5.2 with its published shared attention layout without runtime patches.

What changed

• Update Megatron Bridge to its official GLM-5.2 support commit.
• Update Megatron Core to the exact commit recorded by that Bridge version.
• Refresh the locked dependencies and the generated parallelism reference.

The previous versions created a separate attention indexer on every layer. GLM-5.2 instead creates full indexers on selected layers and shares them with nearby layers. The new pair reads index_topk_freq and index_skip_topk_offset from the official model configuration and builds that layout directly.

Verification

bash format.sh
uv lock --check

A live zai-org/GLM-5.2-FP8 B300 training test is in progress. This PR stays draft until model load, backward calculation, optimizer update, checkpoint reload, and sampling after weight sync pass.