fix: Unpin the Space's torch -- and undo a mis-targeted revert

Two things, because the first attempt at this got it wrong.

The intended change: torch goes back to unconstrained. Pinning <2.9 to force
CUDA 12 wheels did not fix "No CUDA GPUs are available" and was strictly worse.
Unpinned, the log reaches "Waiting for a GPU to become available" then
"Successfully acquired a GPU" before the worker dies, in 14.2s; pinned, neither
line appears and it fails in 1.8s, never reaching GPU acquisition. The CUDA 13
theory was wrong, and the comment now says so, so it is not retried blind.

The correction: 6d404fb claimed to do that but was `git revert HEAD~1`, which on
main was the threading fix, not the pin. It removed USE_INFERENCE_THREAD, the
sync_relay branch and the CLAUDE.md note pointing at 4de16e3 -- re-breaking the
CUDA/thread-context bug this repo has now fixed twice -- while leaving the pin
in place, and it was pushed. Reverted in turn, and the pin removed by editing
the file rather than trusting a commit offset.

Verified after the fact that both survive: on torch neither endpoint touches a
controlai-prefixed thread, on MLX both still run on controlai_0.

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

requirements-space.txt

@@ -9,13 +9,13 @@
 # upper bound here can make the resolve impossible -- pinning transformers <4.56
 # did exactly that, because gradio 6.x needs huggingface-hub >=1.16 and every
 # transformers <4.56 needs <1.0.
-# Upper-bounded on purpose, unlike everything else here. Unconstrained, pip
-# resolved torch 2.11.0, whose wheels carry CUDA 13 (nvidia-cudnn-cu13,
-# cuda-toolkit 13.0.2). ZeroGPU then acquired a GPU and its own forked worker
-# still failed in torch.init(): "RuntimeError: No CUDA GPUs are available".
-# <2.9 keeps torch on CUDA 12 wheels, and on a release `spaces` has had time to
-# patch. Still well inside the platform's own `torch<=2.11.0`.
-torch>=2.4.0,<2.9
+# Left unconstrained deliberately. Pinning <2.9 to force CUDA 12 wheels was
+# tried, on the theory that torch 2.11's CUDA 13 wheels were behind ZeroGPU's
+# "No CUDA GPUs are available". It was not, and the pin was strictly worse:
+# unpinned, the log reaches "Waiting for a GPU" then "Successfully acquired a
+# GPU" before the worker dies (14.2s); pinned, neither line appears at all and
+# it fails in 1.8s, never reaching GPU acquisition. Do not retry that bound.
+torch>=2.4.0
 transformers>=4.51.0
 accelerate>=0.30.0
 bitsandbytes>=0.43.0          # 4-bit NF4; CUDA-only