fix: correct CUDA base image tag to 12.4.1

fix: upgrade to CUDA 12.4 + torch>=2.6 for CVE-2025-32434 compat

fix: pin transformers<5 for google/umt5-base compat (no model_type in config)

fix: remove pre-download step, model downloads at runtime startup

fix: add diffusers, ftfy, regex deps from model dependency tree

fix: add omegaconf and safetensors deps required by model

fix: add lightning dependency required by model

fix: split pip install to avoid index-url conflict for non-torch packages

feat: initial FloodDiffusion streaming demo for HF Space