refactor: RuntimeContext.container_log_dir for every worker-visible log path

[ishandhanani]

Summary

Paths handed to processes inside the container spelled the log mount three different ways: a literal /logs for nsys output, fingerprints, and the Mooncake process config; the CONTAINER_LOG_DIR constant for the measurement window; and the host log_dir for the worker config dump and the torch profiler directory. The host path only works on clusters where the log directory is visible inside the container at the same path, which #431 showed is not every cluster.

RuntimeContext.container_log_dir is the one accessor. It follows the mount from_config creates at CONTAINER_LOG_DIR, so the default stays /logs and a remapped log mount needs no other change. Config dumps, nsys output, the torch profiler directory, fingerprints, the Mooncake process config path, the benchmark profiles directory, and the measurement window directory all read it.

Landing this on main so #431 can drop its _container_log_path helper on rebase; the stacked fork PRs cannot take maintainer pushes because the fork is organization-owned.

Review map

• core/runtime.py: the accessor, and the log mount now uses CONTAINER_LOG_DIR so /logs is defined once.
• cli/mixins/worker_stage.py: eight call sites read the accessor. The torch profiler directory was a host path before.
• cli/mixins/benchmark_stage.py: profiles directory and measurement window directory read the accessor; the constant import goes.
• tests/test_slurm.py: accessor default and remapped behavior, the config dump path both launch paths hand the backend, and a remapped mount driving the dump, profiler, and fingerprint paths. Other test fakes gain the attribute.

Full suite: 2959 passed on both 3.13 and 3.10 locally. About 20 more /logs literals remain in services, frontends, TRT-LLM, exporters, and the nsys observability module for a follow-up.