Native bounding box detection with dynamic text prompting

[Priyadarshini75]

Summary

Adds native bounding box detection directly from SAM3 model outputs and enables
dynamic text prompting at runtime, eliminating the need for hardcoded token IDs.

Changes

Native Bounding Box Detection

• Extract pred_boxes and pred_logits from the TensorRT engine outputs
• Implement CUDA post-processing kernel for bounding box coordinate decoding
• Render green bounding boxes with text labels via OpenCV in the demo app

Dynamic Text Prompting

• Add export_tokenizer.py to export HuggingFace tokenizer files for C++ use
• Add tokenize_prompt.py for standalone prompt tokenization
• Accept the text prompt as a CLI argument instead of compile-time constants
• Dynamically tokenize and feed prompts into the engine at runtime

Dynamic Image Size Processing

• Support variable input image resolutions instead of fixed dimensions
• Improve benchmarking workflow with proper timing isolation

Updated Documentation

• Added usage instructions for bounding box visualization and benchmarking
• Updated README with tokenizer setup steps

Usage

# Export tokenizer (one-time setup)
python3 python/export_tokenizer.py

# Run with any text prompt
./sam3_pcs_app /workspace/test_images /workspace/sam3_fp16.plan "helmet"

# Benchmark mode (no visualization overhead)
./sam3_pcs_app /workspace/test_images /workspace/sam3_fp16.plan "helmet" 1