[UX] Make `infer_split_merge.py --help` work without requiring torch/vLLM installed

[pi-dal]

Problem

Running python infer_split_merge.py --help currently fails on environments where torch (and/or other heavy deps) are not installed, because infer_split_merge.py
imports torch and str2bool inside main() before argparse parses args.

This makes it harder for new users to discover parameters/usage without first setting up a full GPU environment.

Proposed Fix

• Move heavy imports (e.g., import torch) to after args = parser.parse_args().
• Optionally keep a small helper so users can view CLI docs even without full deps.

Notes

Normal execution still requires torch/vLLM; this change is only to improve CLI discoverability.