[Feature] Support lmmseval in unfied eval script

[oscarqjh]

Summary

• Add lmms-eval as a second evaluation backend alongside VLMEvalKit, behind a shared adapter interface
• Refactor run_easi_eval.py from a VLMEvalKit-only script into a backend-agnostic orchestrator
• Users can now run all EASI benchmarks with either backend using aligned CLI options

Changes

Backend adapter pattern (scripts/submissions/backends/):

• base.py — BackendAdapter ABC defining the shared interface (command building, progress polling, score extraction, result archiving)
• vlmevalkit.py — VLMEvalKit adapter extracted from the original run_easi_eval.py
• lmmseval.py — New lmms-eval adapter with task mapping, metric extraction, SiteBench JSONL merge, and resume logic

Orchestrator refactor (run_easi_eval.py):

• New CLI flags: --backend {vlmevalkit,lmms-eval}, --model-args, --accelerate/--no-accelerate, --rerun
• Backend-agnostic benchmark registry (user-facing keys mapped to backend task names by adapters)
• Resume logic: skips completed benchmarks on rerun, --rerun forces fresh evaluation
• ProgressDisplay adapted for lmms-eval (single status icon, no sample-level progress bars)
• Benchmarks run individually per subprocess to avoid HuggingFace datasets filelock deadlocks

Postprocess (postprocess.py):

• build_payload() and build_results_archive() accept optional backend_adapter parameter
• Result archive includes only the latest files per task (no stale results from previous runs)

Environment & docs:

• scripts/setup.sh installs both backends with pinned filelock==3.20.3 and datasets==4.5.0
• Updated docs/Submit_results.md with lmms-eval examples, task name mapping, and resume/rerun docs
• Updated README.md and README_CN.md with unified eval script examples
• Updated lmms-eval submodule to latest

Usage

# VLMEvalKit (default, unchanged)
python scripts/submissions/run_easi_eval.py \
  --model Qwen/Qwen2.5-VL-7B-Instruct --nproc 4

# lmms-eval
python scripts/submissions/run_easi_eval.py \
  --backend lmms-eval \
  --model qwen3_vl \
  --model-args "pretrained=Qwen/Qwen3-VL-8B-Instruct,attn_implementation=flash_attention_2" \
  --nproc 4

# OOM recovery (single-GPU, resumes from where it left off)
python scripts/submissions/run_easi_eval.py \
  --backend lmms-eval --model qwen3_vl \
  --model-args "pretrained=Qwen/Qwen3-VL-8B-Instruct,attn_implementation=flash_attention_2" \
  --no-accelerate

Test plan

• VLMEvalKit path: CLI parsing, command construction, benchmark selection verified via dry-run
• lmms-eval path: full EASI-8 evaluation completed on Qwen3-VL-8B-Instruct with 4 GPUs
• Score extraction: all 8 benchmarks correctly scaled to 0-100 with sub-scores
• Resume logic: completed benchmarks skipped on rerun, --rerun forces fresh evaluation
• Result archive: only latest files per task included in zip
• Submission: successfully submitted to EASI leaderboard