deps: allow vllm to coexist with cu128/cu130 (torch 2.10.0) groups

The Dockerfile always syncs --group=<cuda> --group=vllm, but vllm was in the
same [tool.uv] conflict set as every cuda group, so 'uv sync' rejected the
combination. vllm pins torch==2.10.0, which the cu128/cu130 torch 2.10.0+cuXXX
local builds satisfy; only the cu130-torch213 groups (torch 2.13.0) are truly
incompatible. Move vllm into a conflict set with just the torch-2.13 groups so
cu128+vllm and cu130+vllm resolve. Verified with 'uv sync --locked --dry-run'.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: foreverlms

pyproject.toml

@@ -316,14 +316,21 @@ repository = "https://github.com/NVIDIA/cosmos-framework"
 required-version = ">=0.11.3"
 conflicts = [
   [
-    {group = "vllm"},
     {group = "cu128"},
     {group = "cu128-train"},
     {group = "cu130"},
     {group = "cu130-train"},
     {group = "cu130-torch213"},
     {group = "cu130-torch213-train"},
   ],
+  # vllm pins torch==2.10.0, which is satisfied by the cu128/cu130 torch 2.10.0
+  # local builds, so vllm may combine with those. It is incompatible only with
+  # the torch 2.13.0 groups; keep those mutually exclusive with vllm.
+  [
+    {group = "vllm"},
+    {group = "cu130-torch213"},
+    {group = "cu130-torch213-train"},
+  ],
 ]
 override-dependencies = [
   # Lower bound defeats robomimic/robosuite's old numpy 1.x pins; upper bound