fix(deps): make torch/ML stack opt-in instead of a mandatory install

requirements-amd.txt/requirements-nvidia.txt were pulling a multi-GB
torch wheel by default even though nothing in synapse/ imports torch,
transformers, accelerate, bitsandbytes, or PySide6 - dead weight that
made the pip batch fragile (one failed download could take unrelated
base deps down with it on a slow connection). Split the unused ML/GUI
stack out of requirements-base.txt into a new opt-in requirements-ml.txt,
and dropped the torch lines from the AMD/NVIDIA overlays and generator.

Also: recreate the venv if it exists but pip is missing, instead of
silently reusing a half-built one (ensurepip can fail during venv
creation and leave an interpreter with no pip).

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

requirements-ml.txt

@@ -0,0 +1,20 @@
+# --- Optional: local ML inference stack ---
+# Nothing in synapse/ imports any of this — Ollama handles all inference over
+# HTTP. Only install this if you're doing local model work outside Ollama
+# (fine-tuning, direct transformers inference, etc). Skipping it is what keeps
+# the default install light and out of a multi-GB torch download.
+#
+# This file has no --index-url of its own, so torch resolves as CPU-only
+# unless you combine it with the GPU overlay that sets one:
+#   pip install -r requirements-amd.txt    -r requirements-ml.txt   # ROCm
+#   pip install -r requirements-nvidia.txt -r requirements-ml.txt   # CUDA
+#   pip install -r requirements-ml.txt                              # CPU only
+
+transformers
+accelerate
+bitsandbytes
+safetensors
+sentencepiece
+torch
+torchaudio
+torchvision