[Feature] Switch to python runner for single GPU

[xu-song]

Motivation

There is no need to use torchrun for single GPU inference.
Besides, python runner is more friendly for debugging.
The debugging snippet is as follows

opencompass/opencompass/runners/local.py

Lines 116 to 131 in 889e7e1

	# run in subprocess if starts with torchrun etc.
	if 'python3 ' in cmd or 'python ' in cmd:
	# If it is an infer type task do not reload if
	# the current model has already been loaded.
	if 'infer' in self.task_cfg.type.lower():
	# If a model instance already exists,
	# do not reload it.
	task.run(cur_model=getattr(self, 'cur_model',
	None),
	cur_model_abbr=getattr(
	self, 'cur_model_abbr', None))
	self.cur_model = task.model
	self.cur_model_abbr = model_abbr_from_cfg(
	task.model_cfg)
	else:
	task.run()

[MaiziXiao]

LGTM