[Nanmur]

This change prevents CombineParallelMatmul from throwing when a matched matmul has a non-Var left operand.

The pass groups parallel branches by VarNode identity, so a constant left operand cannot participate in the existing grouping scheme. Skip these ineligible matches and leave the original matmul unchanged.

Fixes #20821

Tests:

• python -m pytest tests/python/relax/test_transform_combine_parallel_matmul.py -q (13 passed)
• pre-commit run --files src/relax/transform/combine_parallel_matmul.cc tests/python/relax/test_transform_combine_parallel_matmul.py
• Windows source build with MSVC 19.51 and Ninja

The issue investigation and patch preparation were assisted by OpenAI Codex.

[Nanmur]

@tlopex Could you please help rerun the Windows check when the runner/setup issue is resolved?

The Windows job has failed twice before any TVM build or test step ran (Actions runs 40280664426 and 40284167740). In both attempts, conda-incubator/setup-miniconda@v4.0.1 failed during environment setup:

1. conda init --all reported needs sudo C:\Users\runneradmin\miniconda3\Scripts\conda.exe.
2. The action retried installation into the same path.
3. The retry exited because C:\Users\runneradmin\miniconda3 was already non-empty.

The subsequent conda info, dependency installation, build, and test steps were skipped, so this failure did not exercise the PR changes. macOS, lint, and all Jenkins checks (arm/cpu/docker/gpu/wasm) are passing. I also cannot rerun the GitHub Actions job from a fork contributor account because it requires repository admin permission.