feat(weight-sync): add ModelExpress trainer publishing

[KavinKrishnan]

Overview

Adds trainer-side publication for a ModelExpress (MX) weight-refit backend. NeMo RL's WeightSynchronizer remains the owner of the refit lifecycle; this PR adds the policy operation that a future ModelExpressWeightSynchronizer will call.

The existing DTensor worker and the Megatron worker publish rank-local tensor ownership and topology metadata. Generation-side discovery, transfer, resharding, and installation remain separate reviews.

Changes

• Adds publish_weights_for_model_express() to the policy interface and trainer workers.
• Publishes native Tensor Parallelism (TP), Pipeline Parallelism (PP), and Expert Parallelism (EP) shards without gathering the full model on rank zero.
• Describes global tensor geometry, rank coordinates, expert ownership, and Megatron-to-Hugging Face name mappings.
• Requires the public ModelExpress publisher lifecycle; no private MX state is accessed.
• Fails before publication for unsupported DTensor placements or unknown Megatron placement at TP greater than one.
• Adds focused lifecycle, geometry, role-classification, and failure tests.
• Adds the trainer-publication design and current implementation status to the documentation.

Architecture boundary

WeightSynchronizer                              [follow-up]
  -> policy.publish_weights_for_model_express() [this PR]
  -> generation discovery and transfer          [follow-up]
  -> inference reshard and install               [follow-up]
  -> global completion                           [follow-up]

This PR does not create a second orchestration path. Existing IPC, HTTP, and NCCL synchronizers are unchanged.

Scope

This PR does not add:

• a user-facing ModelExpress refit configuration;
• generation-worker discovery, transfer, or installation;
• DTensorPolicyWorkerV2 / AutoModel publication;
• Dynamo integration;
• Kubernetes resources, recipes, or benchmark harnesses.

The future synchronizer will own user configuration and construct the internal publisher settings explicitly.

Current status

Partially aligned: rank-local publication is implemented and unit-tested for the existing DTensor worker and the Megatron worker. End-to-end refit requires the generation-worker adapter and the corresponding ModelExpress publisher, planning, and installer APIs.

ModelExpress dependencies

• filtered and versioned discovery: feat(server): add filtered versioned source discovery ai-dynamo/modelexpress#519 (merged);
• mapped vLLM installation: feat(vllm): add mapped direct load installer ai-dynamo/modelexpress#521;
• no-gather geometry and reshard planning: feat(reshard-refit): add no-gather slice-resharding weight refit ai-dynamo/modelexpress#496;
• required publisher contract: public reset_tensors() plus one add_tensor() surface for DTensor and Megatron metadata (follow-up review).

Relationship to #3068

This is the trainer-publication portion of #3068, separated by ownership and review scope. The umbrella remains a reference for the complete integration.

Test plan

• 14 focused publication lifecycle, DTensor geometry, and Megatron classification tests
• Ruff lint and formatting
• Focused pyrefly check for new modules
• Python compilation and Git whitespace checks
• DCO sign-off
• Full repository and GPU end-to-end tests through maintainer CI

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[KavinKrishnan]

/ok to test 5fc0c0b

[KavinKrishnan]

/ok to test 858bf2f