Sliced-only keys on a prep step config crash at workflow construction

[wanghan-iapcm]

dargs accepts template_slice_config, continue_on_num_success and continue_on_success_ratio on a prep step config, but the superops only sanitise run_config before spreading. Setting any of them under a prep step config therefore crashes at workflow-construction time, on input the schema declares valid.

Reproduction

Both at e45c147ed261d7c43b4b5bc9ada6e7d271e32c51.

normalize accepts both keys on a prep config:

template_slice_config      -> ACCEPTED, key present: True
continue_on_success_ratio  -> ACCEPTED, key present: True

(a) template_slice_config — only run_config ever pops it, so it survives into the Step(...) spread:

PrepRunLmp("x", PrepLmp, RunLmp,
           prep_config=normalize({"template_slice_config": {"group_size": 2}}),
           run_config=normalize({}))

TypeError: Step.__init__() got an unexpected keyword argument 'template_slice_config'
  at dpgen2/superop/prep_run_lmp.py:151

(b) continue_on_success_ratio on a non-sliced step. dflow only assigns total inside the dflow_nslices / with_param / with_sequence branches, then uses it unconditionally:

dpgen2/dpgen2/superop/prep_run_lmp.py

Lines 151 to 166 in e45c147

	prep_lmp = Step(
	"prep-lmp",
	template=PythonOPTemplate(
	prep_op,
	output_artifact_archive={"task_paths": None},
	python_packages=upload_python_packages,
	**prep_template_config,
	),
	parameters={
	"lmp_task_grp": prep_run_steps.inputs.parameters["expl_task_grp"],
	},
	artifacts={},
	key=step_keys["prep-lmp"],
	executor=prep_executor,
	**prep_config,
	)

Isolated against dflow directly, a non-sliced step with the key set:

UnboundLocalError: cannot access local variable 'total' where it is not associated with a value
  at dflow/step.py:970

while the same step with with_param=[1, 2, 3] constructs fine — confirming it is specifically the non-sliced case.

Scope

Every prep_run_*.py pops template_slice_config from run_config only, then spreads the unsanitised prep_config into a non-sliced prep step:

file	pops from run_config	spreads **prep_config into a non-sliced step
prep_run_calypso.py	L165	L181
prep_run_lmp.py	L149	L165
prep_run_fp.py	L143	L162
prep_run_dp_train.py	L166	L183

prep_run_diffcsp.py pops at L136 but spreads nothing at all — see the companion issue.

The fix already exists elsewhere

60c8f94 ("fix: continue_on_num_success/continue_on_success_ratio key error", #226, 2024-06-03) solved exactly this for caly_evo_step.py by deriving a sanitised copy before spreading into non-sliced steps:

dpgen2/dpgen2/superop/caly_evo_step.py

Lines 148 to 153 in e45c147

	template_slice_config = run_config.pop("template_slice_config", {})
	expl_mode = caly_evo_step_steps.expl_mode
	no_slice_run_config = deepcopy(run_config)
	no_slice_run_config.pop("continue_on_num_success", None)
	no_slice_run_config.pop("continue_on_success_ratio", None)

The prep side of the four prep_run_*.py files never got the same treatment.

Suggested fix

Give prep_config the same handling run_config gets — pop template_slice_config, and drop the sliced-only continuation keys before spreading into a non-sliced step. Alternatively, reject these keys for prep step configs at the dargs layer so the failure happens at submit-time normalization with a clear message rather than as a TypeError / UnboundLocalError from inside dflow.

Found while reviewing #380.