feat: Move the Space from CPU-only Docker to ZeroGPU (Gradio SDK)

CPU inference on the free cpu-basic tier turned out to be genuinely too
slow for this workload (6.4k-token system prompt + tool schemas before
generation even starts) -- live testing on the Space showed a single
request taking 8+ minutes with zero tokens streamed back. No amount of
context-window or build-flag tuning fixes a fundamentally CPU-bound
problem.

Switches to ZeroGPU: drops the Docker SDK and llama-cpp-python/GGUF
path for the Space, and loads the real fine-tuned model (now uploaded
in full-precision safetensors) via transformers on a dynamically
allocated GPU inside @spaces.GPU-wrapped calls. A minimal Gradio Blocks
app is mounted at /gradio solely to satisfy the platform's Gradio SDK
requirement for ZeroGPU hardware -- the existing custom FastAPI routes
and web UI are untouched and still serve the real app at "/".

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Dockerfile

@@ -1,53 +0,0 @@
-FROM python:3.11-slim
-
-# Prevent interactive prompts & buffer logs
-ENV PYTHONDONTWRITEBYTECODE=1 \
-    PYTHONUNBUFFERED=1 \
-    DEBIAN_FRONTEND=noninteractive \
-    PORT=7860 \
-    HOME=/home/user
-
-# Install essential scientific computation libraries and compilers
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential \
-    cmake \
-    git \
-    curl \
-    liblapack-dev \
-    libblas-dev \
-    gfortran \
-    libopenblas-dev \
-    && rm -rf /var/lib/apt/lists/*
-
-# Set up user for Hugging Face Spaces security standard
-RUN useradd -m -u 1000 user
-WORKDIR /home/user/app
-
-# Install Python dependencies. llama-cpp-python compiles from source (no
-# prebuilt wheel is published for recent versions); cap build parallelism so
-# the compiler doesn't spawn enough jobs to OOM-kill the Spaces build machine.
-# Pin to a fixed AVX2+FMA target instead of NATIVE (avoids depending on the
-# build machine's exact CPU) or full auto-detection (which was the extra
-# compile cost that caused the OOM) -- AVX2/FMA are present on essentially
-# every x86-64 server CPU Spaces runs on, so this keeps inference fast
-# without the crash risk of a mismatched AVX-512 build.
-COPY requirements.txt .
-ENV CMAKE_BUILD_PARALLEL_LEVEL=1 \
-    CMAKE_ARGS="-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_FMA=ON -DGGML_AVX512=OFF"
-RUN pip install --no-cache-dir --upgrade pip && \
-    pip install --no-cache-dir -r requirements.txt
-
-# Copy application files
-COPY --chown=user:user . .
-
-# Ensure storage directories exist with write access
-RUN mkdir -p outputs/plots data/rag_index data/user_docs && \
-    chown -R user:user /home/user
-
-USER user
-
-# Expose standard Hugging Face Spaces port
-EXPOSE 7860
-
-# Launch FastAPI web console
-CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]