[WIP] fix(quickstart): specify model precision before engine building

[yenhao-huang]

IntroNotebooks still uses BuilderFlag.FP16, which raises AttributeError on TensorRT 11, and its PyTorch notebook exports FP32 before requesting FP16 through removed builder options. Migrate the helper to a strongly typed network and make precision explicit in ONNX. Retain fp16_mode through graph conversion with preserved I/O types; use False for already typed models.

The notebook now selects dtype before export, uses consistent normalization, allocates from engine I/O metadata, rejects mismatched inputs, and compares logits with framework references. Disable TF32 for the FP32 reference comparison. Document explicit conversion's accuracy and performance implications, dependencies, and selective mixed precision.

Validation: 5 GPU/error-path tests passed on each of TensorRT 10.16 and 11.2 (11.2 baseline: 3 failed, 2 passed). Pretrained ResNet50 on TRT 11.2 passed FP32/FP16 comparisons at batch 1 and the notebook's batch 32, including the default conversion path and FP16 versus FP32 reference. Batch-32 top-1 predictions all matched. Executed export, preprocessing, allocation, inference, and cleanup notebook cells; Torch references ran on CPU, and Jupyter timing/Torch CUDA cells were not run as a complete notebook. Black, notebook syntax, and git diff --check passed. Report: docs/howard/4581.md.

Addresses NVIDIA#4581. The separate ONNXClassifierWrapper runtime issue covered by NVIDIA#4612/NVIDIA#4613 is outside this precision migration.