build(deps): add transitive research deps and rust-accelerated transfer

The COINs research code transitively imports a chunk of packages that
were missing from requirements.txt — silently filled in by manual
pip-installs in the local mamba env, but absent from the deployment
image. Each one surfaced as a ModuleNotFoundError at gunicorn boot:

- seaborn — graph_analysis/metrics.py top-level import
- karateclub, gensim, python-louvain — community detection paths
- PyMetis, PyGSP — graph partitioning + signal processing
- compress-pickle — graph_data/serialization.py
- python-Levenshtein, six, decorator — karateclub transitive
- tensorboardX — research code logging hooks
- torch-scatter — torch_geometric op fallbacks (PyG wheel index)

Also bumps numpy from 1.23 to 1.26.* (matches the verified-working
local env; the older pin clashed with karateclub's overly-strict and
runtime-incorrect numpy<1.23 cap), and adds hf_transfer to enable
HF_HUB_ENABLE_HF_TRANSFER=1 for ~3x faster cold-start downloads.

src/backend/requirements.txt

@@ -11,6 +11,8 @@ gunicorn>=21.2
 
 # Checkpoint download from Hugging Face Hub (replaces gdown / Google Drive)
 huggingface_hub>=0.25
+# Rust-accelerated file transfer; activated via HF_HUB_ENABLE_HF_TRANSFER=1
+hf_transfer>=0.1
 
 # PyTorch with CUDA 11.8 (falls back to CPU at runtime if no GPU present)
 --extra-index-url https://download.pytorch.org/whl/cu118
@@ -18,14 +20,19 @@ torch==2.0.1+cu118
 torchvision==0.15.2+cu118
 torchaudio==2.0.2+cu118
 
-# PyTorch Geometric (must match torch version)
+# PyTorch Geometric (must match torch version) and the scatter ops it falls
+# back to.  torch-scatter wheel is published in the PyG index alongside.
+--find-links https://data.pyg.org/whl/torch-2.0.1+cu118.html
 torch-geometric==2.3.1
+torch-scatter==2.1.2+pt20cu118
 
 # Research shared deps
 pytorch-lightning==2.0.4
 hydra-core==1.3.2
 omegaconf==2.3.0
-numpy==1.23
+# Pin to 1.26 — matches the verified-working local env.  Older 1.23 pin
+# clashes with karateclub's overly-strict (and runtime-incorrect) numpy<1.23.
+numpy==1.26.*
 pandas>=2.0
 scipy==1.11.0
 igraph==0.9.11
@@ -36,6 +43,21 @@ imageio==2.31.1
 torchmetrics==0.11.4
 tqdm==4.65.0
 scikit-learn>=1.0
+seaborn>=0.13
+tensorboardX>=2.6
+
+# Graph utilities used by COINs (community detection, partitioning, sampling).
+# karateclub itself is installed --no-deps in the Dockerfile because its
+# pyproject pins numpy<1.23 which we know is unnecessary in practice; its
+# real runtime deps are listed individually here.
+gensim>=4.4
+python-louvain>=0.16
+PyMetis>=2023.1
+PyGSP>=0.6
+compress-pickle>=2.1
+python-Levenshtein>=0.27
+six>=1.16
+decorator>=4.4
 
 # MultiProxAn graph generation
 Pillow>=9.5.0