perf: concurrency improvements for high-volume Excel processing

feat: deploy docling first parser

feat: switch to unsloth gguf runtime

feat: switch parser to qwen vl

feat: deploy paddleocr gemini parser

feat: v5.0.0 PaddleOCR-VL-1.5 + Gemini hybrid architecture

fix: increase max-model-len to 65536 for VLM image tokens, improve error logging

fix: total_mem -> total_memory, clean up debug CMD and start.sh

debug: add aggressive startup logging to diagnose silent hang

fix: remove sed pipe that caused block buffering and wrong PID tracking

fix: use HF repo ID for vLLM model loading instead of local path

fix: use JSON format for --limit-mm-per-prompt (vLLM compat)

fix: use python3 in start.sh (vLLM base image)

feat: upgrade to Qwen3-VL-30B-A3B, simplify auth, fix redirects

feat: hybrid VLM parser with Qwen3-VL-8B via vLLM (v2.0.0)