docs: describe training launches in terms of primus-cli

[WangLingxun]

Summary

Documentation-only. Rewrites every guide, example README and config header comment that still told users to export EXP and run examples/run_*.sh, so the whole repository describes launches in terms of runner/primus-cli.

No behaviour changes and no scripts deleted here. examples/run_pretrain.sh and friends still exist on this branch and still work; this PR only stops the documentation from presenting them as the way to launch training. Retiring the scripts, migrating their downstream callers, and the fixes that came out of that migration are in the stacked follow-up #999, which is based on this branch.

• Launch snippets rewritten to the three runner modes — primus-cli direct, primus-cli container, primus-cli slurm srun — across docs/, examples/, and the header comments of the diffusion and SFT configs. Container environment options stay at the container layer rather than leaking past the -- into the Primus CLI.
• examples/README.md gains a mode table explaining when to use direct / container / slurm, the -- argument shape, and where the environment defaults and setup hooks live (runner/helpers/envs/, runner/helpers/hooks/, runner/.primus.yaml).
• References repointed at the code that owns the behaviour instead of at the launcher script: the HipBLASLt tuning master switch and its stage variable at runner/helpers/hooks/train/pretrain/prepare_experiment.sh, deterministic mode at runner/helpers/hooks/05_deterministic.sh and runner/helpers/envs/base_env.sh, and ANP / NCCL_NET_PLUGIN selection at runner/helpers/hooks/03_enable_ainic.sh. The HipBLASLt section previously omitted that the stage is ignored unless PRIMUS_HIPBLASLT_TUNING=1; that is now stated where users hit it.
• New guard test tests/runner/test_docs_cli_examples.py fails when a Markdown command invokes a .sh that is not in the tree, so the next launcher move cannot leave stale instructions behind. It matches only bash / sh / source invocations, so prose that merely names a script stays out of scope. A second assertion pins the hybrid-models guide's --volume / --env DATA_PATH placement, which is the shape that is easy to get wrong when translating a snippet.

Scope boundary

Documentation whose statements only become true once the launchers are actually gone is deliberately not in this PR — it lives in #999 alongside the code that makes it true:

• the Kubernetes section removal, which belongs with the deletion of examples/run_k8s_pretrain.sh;
• the multi-node benchmark commands that name runner/helpers/launch/slurm_pretrain.sh, a helper refactor(runner): retire the examples launchers in favour of primus-cli #999 creates;
• the auto-benchmark and ODC READMEs, whose scripts refactor(runner): retire the examples launchers in favour of primus-cli #999 repoints at primus-cli;
• the data-preparation guide's TOKENIZED_*_DATA_PATH / split-seed / rank-coordination sections, which describe prepare-hook behaviour refactor(runner): retire the examples launchers in favour of primus-cli #999 restores.

The rule applied throughout: this branch must be self-consistent on its own. Nothing here references a file that does not exist yet, and nothing contradicts the tree as it stands.

Test plan

• pre-commit run --from-ref origin/main --to-ref HEAD — clean.
• tests/runner/run_all_tests.sh — 10/10 suites.
• pytest tests/runner/ — 2 passed, including test_docs_cli_examples.py, which confirms every documented script invocation resolves against this branch's tree (the old examples/run_*.sh scripts are still present here and still referenced where they have not been rewritten).
• Stacking verified: this branch plus refactor(runner): retire the examples launchers in favour of primus-cli #999 reproduces the pre-split tree byte for byte (identical tree hash), so splitting lost nothing.