feat: [draft do not merge] Random dataset with specified input and output sequence length

[guyueh1]

What does this PR do ?

Random dataset following specified input and output sequence length

Issues

closes #1302

Usage

Use the following flags for fixed ISL/OSL eval

uv run examples/run_eval_random_dataset.py \
+data.input_len_or_input_len_generator=1000 \
generation.ignore_eos=true \
generation.vllm_cfg.max_model_len=3000

Use the following flags for fixed ISL/OSL GRPO

uv run examples/run_grpo_random_dataset.py \
+data.input_len_or_input_len_generator=1000 \
policy.generation.ignore_eos=true \
policy.generation.output_len_or_output_len_generator=2000 

Use the following flags for random ISL/OSL GRPO with mean + stdv

uv run examples/run_grpo_random_dataset.py \
grpo.val_at_start=false \
grpo.val_period=0 \
policy.max_total_sequence_length=8000 \
+data.input_len_or_input_len_generator.mean=1000 \
+data.input_len_or_input_len_generator.std=100 \
+policy.generation.output_len_or_output_len_generator.mean=2000 \
+policy.generation.output_len_or_output_len_generator.std=1000 \
policy.generation.ignore_eos=True

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

Summary by CodeRabbit

Release Notes

• New Features

  • Added example scripts for GRPO training and evaluation workflows with support for random datasets
  • Introduced ignore_eos flag in generation configurations for flexible EOS token handling
  • Added output length configuration options for generation control
  • Implemented dummy environment for testing and evaluation scenarios
  • Added FP8 quantization support for MoE (Mixture of Experts) modules

• Refactor

  • Extended generation configuration to support configurable sequence length generation
  • Enhanced model initialization with improved timing instrumentation for evaluation workflows