Add --num_repeats to flexeval_lm command

[butsugiri]

This PR adds a new argument --num_repeats to flexeval_lm command.

What this PR does

For the given command...

flexeval_lm \
  --language_model HuggingFaceLM \
  --language_model.model "sbintuitions/tiny-lm" \
  --eval_setups="./tests/dummy_modules/configs/eval_suite.jsonnet" \
  --save_dir "results-multi" \
  --force=true \
  --num_repeats=3

The final directory structure is as follows:

.
├── generation
│   ├── run0
│   │   ├── config.json
│   │   ├── metrics.json
│   │   └── outputs.jsonl
│   ├── run1
│   │   ├── config.json
│   │   ├── metrics.json
│   │   └── outputs.jsonl
│   └── run2
│       ├── config.json
│       ├── metrics.json
│       └── outputs.jsonl
├── multiple_choice
│   ├── run0
│   │   ├── config.json
│   │   ├── metrics.json
│   │   └── outputs.jsonl
│   ├── run1
│   │   ├── config.json
│   │   ├── metrics.json
│   │   └── outputs.jsonl
│   └── run2
│       ├── config.json
│       ├── metrics.json
│       └── outputs.jsonl
└── perplexity
    ├── run0
    │   ├── config.json
    │   └── metrics.json
    ├── run1
    │   ├── config.json
    │   └── metrics.json
    └── run2
        ├── config.json
        └── metrics.json

What this PR does not

• Aggregating the result from multiple generations
  • I am wondering if we should include this feature in this PR

[junya-takayama]

Nice feature👏
LGTM!