[CI] Fix 4-GPU test timeout by using 3 partitions

[alisonshao]

Summary

Fixes the unit-test-backend-4-gpu timeout issue introduced by #14222.

PR #14222 moved test_piecewise_cuda_graph.py (estimated 1200s) to the per-commit-4-gpu suite, which caused the LPT partition algorithm to create an unbalanced distribution with only 2 partitions:

• Partition 0: 1264s (2 tests) ✓
• Partition 1: 1483s (4 tests) ✗ exceeds 20min timeout

This PR increases the number of partitions from 2 to 3:

• Partition 0: ~1200s (1 test) - test_piecewise_cuda_graph.py
• Partition 1: ~772s (2 tests) - test_pp_single_node.py, test_qwen3_next_models.py
• Partition 2: ~775s (3 tests) - test_local_attn.py, test_gpt_oss_4gpu.py, test_multi_instance_release_memory_occupation.py

All partitions now fit within the 20-minute timeout.

Example failure

https://github.com/sgl-project/sglang/actions/runs/19845982270/job/56878342316?pr=14253

args=Namespace(timeout_per_file=1200, suite='per-commit-4-gpu', auto_partition_id=1, auto_partition_size=2, continue_on_error=False)
The running tests are  ['test_pp_single_node.py', 'test_local_attn.py', 'test_gpt_oss_4gpu.py', 'models/test_qwen3_next_models.py']
...
Error: The action 'Run test' has timed out after 20 minutes.

Test plan

• CI passes with the new 3-partition configuration

[alisonshao]

/tag-and-rerun-ci