Add sp_min_token_num=0 to E2E correctness tests for SP and AsyncTP

[Copilot]

When sp_min_token_num is unset, vLLM auto-computes an SP threshold via get_sequence_parallelism_threshold(). For small models (e.g. hmellor/tiny-random-LlamaForCausalLM) or dummy-loaded models with overridden hidden_size, this returns None (below SP_MIN_HIDDEN_SIZE), silently disabling SP — meaning the correctness tests weren't actually exercising SP or AsyncTP paths.

Changes

• test_sequence_parallel.py: Add "sp_min_token_num": 0 to pass_config in _compare_sp() to force SP enabled regardless of model size.
• test_async_tp.py: Add "sp_min_token_num": 0 to pass_config in test_async_tp_pass_correctness() — AsyncTP internally requires SP, same silent-disable risk applies.

Consistent with the existing pattern in tests/compile/fusions_e2e/test_tp2_async_tp.py which already sets sp_min_token_num=512 for the same reason. Using 0 is more aggressive: SP applies at any batch size.

compilation_config = {
    "mode": CompilationMode.VLLM_COMPILE,
    "compile_sizes": [4, 8],
    "pass_config": {
        "enable_sp": True,
        "fuse_gemm_comms": fuse_gemm_comms,
        # Override threshold to ensure SP is applied even for small models.
        "sp_min_token_num": 0,
    },
}

---

⚡ Quickly spin up Copilot coding agent tasks from anywhere on your macOS or Windows machine with Raycast.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Copilot.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[WoosukKwon]

Closing as superseded by #60517, which removed the torch.compile-based sequence parallelism pass and AsyncTP (along with enable_sp,
fuse_gemm_comms, sp_min_token_num, and their tests). The code this PR modifies no longer exists. Thanks for the contribution!