refactor: remove devices related gflags.

[XuZhang99]

Remove the --devices, --device_id, and --draft_devices CLI flags along with their c_api/cc_api/pybind surfaces. Device selection now defaults to "auto": each process enumerates its visible devices via Platform::device_count(), honoring *_VISIBLE_DEVICES masks. Add a unified Platform::current_device() (NPU/CUDA/MLU/DCU/MUSA).

Rework DistManager::setup_multi_node_workers for multi-process serving: the parsed devices vector is now the visible-card pool rather than "cards this process uses". Each process runs exactly one worker bound to devices[node_rank], so world_size == nnodes instead of devices.size()*nnodes. This keeps every process able to see all cards (so ATB can build the cross-card HCCL comm group) while deriving worker cardinality from nnodes/node_rank. Verified: 2-process TP=2 Qwen3-8B serves correctly, gsm8k mean_acc=0.9234.

Update launch docs/examples to select devices via *_VISIBLE_DEVICES instead of --devices.

Description

Related Issues

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

Thank you for contributing to xLLM. Before requesting review, please make sure the following items are complete.

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Allowed types: feat, bugfix, docs, test, refactor, chore, style, revert, perf, model, build, release.
The subject should use clear English, start with a verb, include at least 4 words, and end with ..

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

If you are unsure how to set up pre-commit, see the pre-commit documentation.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

[pjgao]

在 PR1929 的实际合入提交上叠加了 PR1948 的两个提交，并完成 Qwen3.5-2B MTP 冒烟和性能验证。

测试条件

• Target TP2，Draft Body TP1，Draft LMHead TP2
• graph 开启，graph double buffer 关闭，prefix cache 关闭
• 输出 25 token
• warmup 8 条后，20 条请求执行 3 轮，共统计 60 条
• MTP4、MTP5 分别完整执行两次

功能验证

• PR1948 可以正确从可见设备中自动发现并映射两个 rank
• Draft Body TP1 配置正常生效
• 每轮均为 60/60 请求成功，没有启动、推理或退出异常

服务端性能结果

配置	PR1929 基线 E2E / TTFT / TPOT	叠加 PR1948 两轮均值 E2E / TTFT / TPOT
MTP4	152.940 / 78.500 / 3.105 ms	154.801 / 78.983 / 3.156 ms
MTP5	154.497 / 80.017 / 3.098 ms	156.798 / 79.608 / 3.216 ms

叠加后的数值略慢，但两次重复间存在明显波动，尤其 MTP5 两次 E2E 分别为 158.498 ms 和 155.098 ms。PR1948 本身没有修改 draft、verify 或 decode 热路径，因此估计这里是测试波动。

仅从 PR1929 相关功能兼容性看，应该是没问题的。