[Fix] Correct LOCAL_RANK default value to 0 in run.py

[pranaysb]

Fixes #1589

Description

In run.py, LOCAL_RANK was being read from the environment with an incorrect default value of 1. If LOCAL_WORLD_SIZE is set but LOCAL_RANK is not, this default causes the GPU assignment block to skip the first GPU partition (DEVICE_START_IDX = GPU_PER_PROC * 1).

This PR corrects the default value of LOCAL_RANK to 0.

Root Cause & Justification

• Conventions: PyTorch torchrun conventions state that LOCAL_RANK always starts at 0.
• Consistency: The RANK variable in run.py:33 correctly defaults to 0, and LOCAL_RANK shares the exact same semantic concept.
• Codebase Alignment: Every other LOCAL_RANK read across this codebase correctly defaults to 0 (see vlmeval/vlm/emu.py:76, vlmeval/vlm/parrot.py:61, and vlmeval/vlm/parrot.py:188).

Verification

• pre-commit run --all-files passed successfully.
• flake8 checks on run.py passed cleanly (ignoring W503 as per setup.cfg / .pre-commit-config.yaml).
• Verified git diff shows exactly one line changed.