feat(deployment): refactor config and clean code

[vividf]

Summary

This refactoring introduces a unified component configuration system to simplify deployment configuration for multi-component models (e.g., CenterPoint) and cleans up API interfaces to improve maintainability and type safety.

Key Changes

1. Unified Component Configuration System

Problem: Configuration was scattered across multiple places (model_io, backend_config, onnx_config, etc.), making it difficult to configure multi-component models and maintain consistency.

Solution: Introduce a unified components configuration structure as a single source of truth. Each component defines its name, file paths, I/O specification, and TensorRT profile in one place.

2. Improved Artifact Path Resolution

Problem: Path resolution logic was scattered with no unified interface, missing support for multi-component models and path validation.

Solution: Add centralized path resolution functions with unified strategy. Require base_dir to be a directory and read filenames from components config. Add path security checks to prevent traversal and validate file existence.

####3. Configuration Class Refactoring

• Rename BackendConfig → TensorRTConfig for clearer naming
• Simplify TensorRT config structure by removing nested common_config
• Move TensorRT profiles to components.*.tensorrt_profile for per-component configuration
• Update config accessors to read from unified components structure

4. Evaluation System Improvements

• Support multi-evaluator mode with metric prefixes to avoid key conflicts
• Update DetectionSummary to support multiple matching modes
• Improve result display format for mAP_by_mode and mAPH_by_mode

** Note: I will create a separate PR to refactor the detection3d metrics after completing the t4metrics refactor. At that point, detection3d will directly use t4metricsv2.**