Fix onnxruntime-gpu churn in Docker TensorRT FP8 install step

[forkni]

Summary

• nvidia-modelopt[onnx] was unbounded here, floating to 0.45.0 — whose [onnx] extra force-upgrades onnx to 1.21.0 (breaks FP8 quant: external-data loading -> negative QDQ scale) and hard-pins onnxruntime-gpu==1.22.0 on Windows, downgrading off the onnx==1.19.1/onnxruntime-gpu==1.24.4 pins installed just above.
• Pin nvidia-modelopt==0.43.0 (the same proven pin already used in forkni/StreamDiffusion-installer) and drop the [onnx] extra, enumerating its deterministic deps explicitly instead. modelopt core has no onnx requirement, so the churn/regression never enters the resolve.
• Mirrors Fix onnxruntime-gpu churn in TensorRT FP8 install step StreamDiffusion-installer#4 and Fix onnxruntime-gpu churn in Docker TensorRT FP8 install step #14 (SDTD_040_beta_release, both already merged), applied here to the same Docker/Linux-only tool on SDTD_032_dev.

Test plan

• py_compile src/streamdiffusion/tools/install-tensorrt.py passes
• Confirmed this file was byte-identical to the pre-patch SDTD_040_beta_release copy before patching, so the same enumeration verified via pip --dry-run on 040 applies unchanged

🤖 Generated with Claude Code

[claude]

Claude finished @forkni's task in 2m 20s —— View job

---

Code review

Summary: This PR pins nvidia-modelopt==0.43.0 and drops its [onnx] extra in src/streamdiffusion/tools/install-tensorrt.py, replacing it with an explicit enumeration of the extra's remaining deps. This avoids the extra pulling in onnx==1.21.0 / downgrading onnxruntime-gpu off the pins installed a few lines above. The change is scoped to a single 13-line block, Docker/Linux-only CUDA-12 install path.

No issues found. Checked for bugs, logic errors, and dependency-resolution problems in the diff.

One candidate concern was investigated and ruled out: cppimport, lief, and ml_dtypes are added unpinned in the new command. However, this is not a regression introduced by this PR — the previous nvidia-modelopt[onnx] already pulled these same transitive deps with the same (absent) version constraints, since pip extras don't add extra pinning beyond what's declared. This PR just makes the resolution explicit; behavior is unchanged.