Make agent delegation be able to be specified in benchmarks

[neubig]

Currently in benchmarks, we don't support sub-agent delegation.

benchmarks/benchmarks/swtbench/run_infer.py

Lines 260 to 265 in 680ce0f

	# TODO: we can enable condenser and security analyzer later
	# and have them configurable via EvalMetadata
	# condenser=get_default_condenser(
	# llm=self.metadata.llm.model_copy(update={"service_id": "condenser"})
	# ),
	# security_analyzer=LLMSecurityAnalyzer(),

We'd like to support this and try, e.g. SWE-Bench with agent delegation on.