Simplify CLI parameters for run-infer / run-eval. Avoid double defaults

[simonrosenberg]

There are default values in evaluation repo for DATASET / DATASET_SPLIT / GAIA_LEVEL / etc ...
And then those values have defaults again in benchmarks repo.
This duplication causes errors since someone must change a value in two very different places to update a parameter.

• the evaluation repository should not define any of those values. It should rely on the default values of the benchmarks repo for run-infer run-eval
• all such values should be in a {benchmarks}/config.py
• all such values should be saved in the artifacts for traceability

Also, SWTBench should have

• DATASET_INFER = eth-sri/SWT-bench_Verified_bm25_27k_zsp
• DATASET_EVAL = princeton-nlp/SWE-bench_Verified