feat(training): adopt CUDA_VISIBLE_DEVICES as the rank-local GPU contract

[TATP-233]

Problem

The current launch topology mixes three user-facing concepts:

1. training.devices (host-visible CUDA indices)
2. rank-local CUDA index remapping inside launcher children
3. backend payload fields and inference/ring/learner device strings

Our required policy is simpler:

• one rank owns exactly one GPU;
• learner, collector public plane, inference ring, replay, and backend payload all share that GPU;
• CPU-physics backends bind each rank to a CPU block;
• no cross-device CUDA IPC inside one rank.

The strict same-device guard in resolve_inference_placement is correct and should remain; the topology source and configuration burden should change.

Proposed contract

Rank-local visibility

• Single-rank execution: an externally set single-entry CUDA_VISIBLE_DEVICES is authoritative. The rank uses cuda:0 in its local namespace; no training.devices is required.
• Multi-rank launch: each rank child receives exactly one visible physical GPU (CUDA_VISIBLE_DEVICES=<one entry>) plus existing rank metadata. All in-rank consumers use cuda:0; remapped global indices never reach EnvCfg/backend payloads.
• Preserve UUID-based parent CUDA_VISIBLE_DEVICES entries when mapping physical devices.

Configuration precedence

1. external rank-local single-entry CUDA_VISIBLE_DEVICES: authoritative;
2. training.devices: compatibility multi-rank selection used to partition/remap visibility before children start;
3. conflict between an in-rank explicit device and visibility fails closed.

training.devices remains supported during transition but is not required for ordinary single-rank training.

CPU-physics rank binding

• dp_collector_cpu_ids / EnvCfg.cpu_ids remains the explicit CPU block owner.
• Automatic partitioning respects process affinity; NUMA/GPU locality remains 多 GPU data-parallel 缺 NUMA/PCIe 拓扑感知：rank→GPU 映射不查 NUMA 亲和性 #1505/NUMA 主机上 collector CPU 默认切片不感知拓扑：需按 GPU NUMA 亲和性自动推导 #1506.
• One CPU-sim rank keeps physics workers and host packed compute inside its block.

Acceptance criteria

• CUDA_VISIBLE_DEVICES=<one GPU> without training.devices produces env/ring/learner all cuda:0.
• Multi-rank children each see exactly one GPU.
• Backend payload fields resolve to local cuda:0.
• Cross-device requests fail closed.
• training.devices remains compatible.
• CPU backend partitioning retains packed HOST_BRIDGE behavior.
• Launcher/placement/backend mapping contract tests pass.

[TATP-233]

Minimal topology PR is up: #1798, with the launcher/uni_rl side in unilabsim/unilab_rl#70. Scope is limited to rank-local visibility and local cuda:0 resolution; owner YAML device semantics, CPU NUMA, and nan-guard work remain separate.

[TATP-233]

Implemented in #1798 and unilabsim/unilab_rl#70. CUDA_VISIBLE_DEVICES is now the rank-local boundary, children receive one GPU each, in-rank consumers resolve to local cuda:0, and conflicts fail closed. training.devices remains compatible as the multi-rank partitioner.