[platform] fix: quiet CPU-only Ray worker warnings

[Wooonster]

What does this PR do?

Downgrade two platform auto-detection warnings to DEBUG only when verl can safely confirm that the current process is a Ray task or actor with no assigned GPU/NPU resources:

• No supported accelerator detected ... Falling back to 'nvidia'.
• Platform 'nvidia' (PlatformCUDA) is registered but not available ...

CPU-only actors such as AgentLoopWorker and RewardLoopWorker intentionally do not request accelerator resources, so these warnings are expected and make healthy distributed RL runs look broken.

The helper is fail-closed:

• Ray remains an optional dependency.
• A missing/uninitialized Ray installation, a driver process, or any runtime-context query failure preserves WARNING behavior.
• Ray workers assigned GPU or NPU resources preserve WARNING behavior, so real accelerator visibility failures remain visible.
• Platform detection, fallback selection, resource allocation, and visible-device environment variables are unchanged.

Root cause

The platform abstraction probes accelerator availability process-locally. Intentional CPU-only Ray actors see no CUDA/NPU device, so both the auto-detection fallback and unavailable-platform paths log at WARNING even though those actors never perform accelerator work.

Duplicate-work check

No open PR was found for issue #6741 or the exact warning text.

• [hardware] feat: add platform abstraction layer and plugin-based engine override system #6086 introduced the platform abstraction where the diagnostics originate.
• 【Trainer】Ray actor 被识别为 CPU-only，导致 NPU 未生效 #6741 is a real NPU allocation/visibility failure and is intentionally not suppressed by this change.
• [trainer] feat: add V1 trainer support for the RemoteBackend plugin abstraction #7102 mentions CPU-only forwarder actors but does not change platform diagnostics.
• [worker, perf] feat: defer scalar metric materialization #7096 only mentions the fallback in CPU test notes and does not fix it.

Test

The human submitter reviewed every changed line and personally ran:

.venv/bin/pytest -q tests/plugin/test_platform_abstraction.py
# 17 passed

.venv/bin/pre-commit run ruff --files \
  verl/plugin/platform/platform_manager.py \
  tests/plugin/test_platform_abstraction.py
# Passed

.venv/bin/pre-commit run ruff-format --files \
  verl/plugin/platform/platform_manager.py \
  tests/plugin/test_platform_abstraction.py
# Passed

Focused mocked-Ray coverage includes:

1. CPU-only Ray worker: both messages are DEBUG and no WARNING is emitted.
2. GPU-assigned Ray worker: both WARNING messages remain.
3. Ray unavailable: existing WARNING behavior remains.
4. Ray runtime-context failure: existing WARNING behavior remains.

AI assistance

OpenAI Codex assisted with repository exploration, implementation, tests, duplicate-work checks, and PR drafting. The human submitter reviewed every changed line, understands the logging-only behavior and fail-closed resource detection, and personally ran the test and formatting commands above.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.