Add ONNX models documentation and benchmark results

Fix: remove outdated MTEB pending note (results already included)

Add full latency comparison table (512-4K tokens, batch 1 & 8)

Add latency comparison vs BGE-M3 and Qwen3-Embedding-0.6B

Add MTEB benchmark results (61.4 mean, 79.3 STS)

Upload mmBERT-Embed-32K-2D-Matryoshka: 80.5 STS, 307M params, 32K context