[Feature]: Expose SmoothQuant alpha for int8_sq quantization in quantize.py

🚀 The feature, motivation and pitch

ModelOpt's SmoothQuant calibration defaults to alpha=1.0 and TensorRT-LLM's
examples/quantization/quantize.py exposes no way to change it — the only override in the
tree is the hardcoded Gemma special case in quantize_by_modelopt.py
({"method": "smoothquant", "alpha": 0.5}). The SmoothQuant paper recommends 0.5–0.9
depending on the model.

In our sweeps on Qwen3-0.6B/1.7B (11 alphas × 2 calib sets, real TRT engines, GSM8K-1319),
alpha=1.0 was never optimal and cost up to 20 GSM8K points vs the best alpha in the same
setting (worst case: Qwen3-0.6B, 15.39% at alpha=1.0 vs 35.41% at alpha=0.85, against a
36.32% bf16 baseline).

Proposal: add --smoothquant_alpha (float, default None = current behavior) to quantize.py
and thread it to quantize_and_export, where it sets the same algorithm dict the Gemma
special case already uses. Backward compatible; ModelOpt already pydantic-validates the
field (bounds [0, 1]). Happy to submit the PR (change is ready).

Alternatives

ModelOpt's hf_ptq.py has a --recipe mechanism, so an alternative is to contribute a recipe
YAML on the Model-Optimizer side instead. But TRT-LLM's quantize.py has no recipe mechanism,
so exposing the flag directly is the minimal, self-contained fix for this CLI. The two
approaches don't conflict.

Additional context

Full sweep (11 alphas × 2 calib sets × 2 Qwen3 sizes):

The PR is ready: +121/−1, includes a GPU-free unit test
(tests/unittest/trt/quantization/test_smoothquant_alpha.py). No behavior change when the
flag is unset.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.