Add LongBench v2 support for DeepSeek

[yiliu30]

This PR adds LongBench v2 support for DeepSeek, similar to the Qwen implementation in PR #2406.

Changes

• DeepSeek LongBench v2 support

Type of Change

Feature enhancement - adds long-context evaluation support for DeepSeek models

Description

Adapts the LongBench v2 evaluation framework for DeepSeek models:

• Dynamic configuration: Automatically adjusts max_length to 40960 for longbench tasks (vs 8192 for standard tasks)
• Server-based evaluation: Uses vLLM server with API-based evaluation for better stability with long contexts
• Modular functions: Refactored code into reusable functions for server management and evaluation
• Parallel execution: Supports 512 threads for efficient longbench evaluation

Implementation Details

1. Added long-bench-eval dependency to requirements.txt
2. Refactored run_evaluation.sh with:
   • Task-based routing (detects 'longbench' in task name)
   • vLLM server lifecycle management (start, health check, cleanup)
   • Proper error handling and logging
3. Maintains backward compatibility with standard lm-eval tasks

Expected Behavior

When task name contains 'longbench', the script will:

1. Start a vLLM server with 40K context length support
2. Wait for server to be ready (with health checks)
3. Run LongBench evaluation via API
4. Clean up server on completion or error

For standard tasks, uses the original direct lm-eval execution.

How has this PR been tested?

Mirrors the Qwen implementation from PR #2406

Dependency Change

Added: long-bench-eval @ git+https://github.com/yiliu30/long-bench-eval