fix(megatron): forward mtp_shared_weights to megatron-bridge

[Excelius-Wang]

What does this PR do?

Forward mtp_shared_weights to mtp_use_repeated_layer when using the megatron-bridge backend.

The default mcore-bridge backend already consumes the same-name field in mcore_bridge.ModelConfig; this patch covers the provider path, where automatic name matching cannot see the renamed field.

Why is this needed?

With --mtp_num_layers N --mtp_shared_weights true, megatron-bridge previously left mtp_use_repeated_layer=False, so it built independent MTP layers instead of one repeated layer.

The mapping is applied only when the opt-in flag is true, so model-specific provider defaults remain unchanged when it is disabled.

Testing

• Added a CPU/mock regression test.
• Covered opt-in mapping, disabled-flag behavior, and preservation of provider defaults.
• Verified the mapped provider override and model attachment.
• compileall, Flake8, YAPF, and git diff --check pass.

Related to #9936