Install pre-built causal-conv1d wheel at startup for fast inference path

Remove causal-conv1d (OOMKilled during build), will use pre-built wheel

Re-add causal-conv1d, CUDA_HOME env var set for nvcc

Remove causal-conv1d (needs CUDA compilation), keep flash-linear-attention

Add flash-linear-attention + causal-conv1d for fast inference path

Re-enable SSR, keep show_error

Fix markdownify escape() compatibility

Enable show_error and disable SSR for debugging

Add error handling and debug logging to inference

Fix examples: download image locally to avoid SSRF validation

Add torchvision for Qwen3VLVideoProcessor

Upgrade transformers to 5.3.0 for qwen3_5 support

Initial chandra-ocr-2 ZeroGPU Space

initial commit