ci: allocate idle CI devices dynamically

[zhangyue207]

Summary

• Parse Ascend npu-smi info process table entries into process_count and process_pids.
• Parse Iluvatar ixsmi process table entries into process_count and process_pids.
• Exclude Ascend and Iluvatar devices with running processes from automatic allocation and fail allocation when no process-free device is available.
• Remove the platform-level hard-coded /dev/davinci0 mount so selected Ascend devices are mounted dynamically by run.py.
• Increase queued GitHub CI watchdog timeout from 10 minutes to 30 minutes in both CI workflows.

Motivation

CI jobs using automatic allocation could still enter containers on devices with leftover processes. The wrapper should choose a process-free device before container launch and expose process metadata in status output for debugging.

The GitHub queued-job watchdog also needs more time for self-hosted runner scheduling, so this PR extends that watchdog from 600 seconds to 1800 seconds.

Scope

This PR targets codex/prune-unused-ci-artifacts, which contains the .ci submodule commit currently pinned by InfiniOps upstream. This keeps the review diff limited to dynamic device allocation and CI watchdog timeout fixes.

Test Plan

Check	Result
pytest tests/test_resource.py tests/test_run.py tests/test_workflow.py tests/test_shadow_workflow.py -v	98 passed
RUFF_CACHE_DIR=/tmp/infini-ci-ruff-cache ruff check ci_resource.py tests/test_resource.py tests/test_run.py tests/test_workflow.py tests/test_shadow_workflow.py	passed
RUFF_CACHE_DIR=/tmp/infini-ci-ruff-cache ruff format --check ci_resource.py tests/test_resource.py tests/test_run.py tests/test_workflow.py tests/test_shadow_workflow.py	passed
git diff --check	passed
python run.py --config config.yml --job ascend_npu --dry-run --gpu-id 3	emitted --device=/dev/davinci3:/dev/davinci0; no platform-level /dev/davinci0
python run.py --config config.yml --job ascend_npu --dry-run	auto selected NPU 1 on the local Ascend host
python run.py --config config.yml --job iluvatar_gpu --dry-run --gpu-id 2	emitted CUDA_VISIBLE_DEVICES=2

Notes

• Existing Ascend container mapping keeps physical device IDs outside the container and maps the selected device to logical davinci0 inside the container.
• Iluvatar keeps the existing /dev passthrough and CUDA_VISIBLE_DEVICES={id} container behavior; this PR changes only which device is selected automatically.
• Local Iluvatar hardware is not available in this environment, so automatic ixsmi selection is covered by parser and allocation unit tests.
• InfiniOps will need a follow-up submodule pointer update after this CI change is accepted.