feat: TensorRT FP16 backend for depth estimation

[Intersteller-Apex]

Summary

Adds TensorRT FP16 as an inference backend for the depth estimation skill.

3 files changed, 194 insertions, 0 deletions — purely additive, zero existing lines modified.

Changes

• transform.py (+182 lines): TRT engine building, caching, and inference methods
• models.json (+7 lines): TRT FP16 variant entry under win32
• requirements.txt (+5 lines): Optional tensorrt>=10.0 dep (non-macOS)

How it works

Backend selection in load_model() tries each backend in order:

1. CoreML (macOS only)
2. TensorRT — import tensorrt fails fast if not installed
3. PyTorch (fallback)

First run builds the engine (30-120s), subsequent runs load from cache (~instant).
Engine files are GPU-specific: ~/.aegis-ai/models/feature-extraction/trt_engines/

Protocol compliance

Follows the JSONL stdin/stdout protocol. Ready event emits backend: tensorrt.
No changes to transform_base.py, benchmark scripts, or CoreML/PyTorch code paths.

Benchmark (RTX 4070 Laptop)

Backend	Avg Latency	FPS
PyTorch CUDA FP32	36.48ms	27.4
TensorRT FP16	5.29ms	189.0

6.9x speedup

[Intersteller-Apex]

Lightweight changes