Gracefully skip BigBench tasks with no data & guard final aggregation

[NourFahmy]

Description:
Running the harness on BigBench stalls (or crashes) once it reaches bigbench_simple_arithmetic_multiple_targets_json_generate_until because that task’s configured split (“train”) doesn’t exist, causing:

• A ValueError: Instruction "train" corresponds to no data! in datasets.load_dataset().

• An unguarded “Test One Doc” block in ConfigurableTask.init that blindly indexes eval_docs[0].

• Later, an UnboundLocalError in evaluate() when referencing show_group_table if all tasks have zero examples.

This PR makes the harness robust to missing-split/empty tasks by:

• Catching missing-split errors in ConfigurableTask.download() and substituting an empty DatasetDict({"default": …}).

• Wrapping the “one‐doc” sanity checks in ConfigurableTask.init inside if self.eval_docs: (with an else fallback for minimal attribute initialization).

• Filtering out zero-example tasks at the top of evaluate(), so they never build requests or invoke the LM.

• Initializing show_group_table=False before the group-aggregation step and switching from bitwise & to logical and when inserting the "groups" entry in results_dict.

These changes resolve the crash and prevent any lag or endless stalling when the harness reaches that BigBench task, or other tasks that experience similar issues.

[baberabb]

Sorry for the tardiness, I'll review soon. My main hesitation is we do not want tasks with misconfigured configs to silently work, as that would make maintenance much harder: we'd lose visibility into which tasks are actually broken/misconfigured and would be especially problematic for aggregating group scores. I'll have to think what workaround works better. Currently thinking an exclude-tasks param might work so users can explicitly skip some tasks.