fix(test): seed torch._inductor.test_operators to fix test-api shard failures

test-api's reranker-bearing shard (consistently 2/3) has been failing on
every run with a misleading "sentence-transformers is required for
LocalSTEmbeddings" ImportError. The real cause is torch, not the deps:

  import torch._inductor.test_operators
  torch/_inductor/test_operators.py:8:
    _test_lib_def = torch.library.Library("_inductor_test", "DEF")
  RuntimeError: Only a single TORCH_LIBRARY can be used to register the
    namespace _inductor_test ...

transformers' lazy loader imports torch._inductor.test_operators while
resolving AutoModelForSequenceClassification / GenerationMixin (used by the
cross-encoder). Under pytest-xdist that module's body executes twice, and
its module-level `_inductor_test` TORCH_LIBRARY re-registration is rejected
by torch. transformers wraps the RuntimeError and re-raises it as the
sentence-transformers ImportError, so the symptom points at the wrong dep.

This is the same class of bug the adjacent `import torch` seed already
guards against (torch/overrides.py double-init); extend it to the inductor
test-operators module. Seeding it once at conftest import puts it in
sys.modules so the later lazy import is a cache hit and the body never
re-executes. Verified locally that a re-import after seeding is a no-op
(no RuntimeError). Guarded so slim/no-torch envs still collect.

Reproduces independent of dependency versions (seen at transformers 5.3.0
and 5.12.1, torch 2.10 and 2.12), which is why it blocked every recent
uv.lock PR including Dependabot's.

Author: dcbouius

hindsight-api-slim/tests/conftest.py

@@ -24,6 +24,19 @@
 # per worker process. Guarded so slim/no-torch environments still collect.
 try:
     import torch  # noqa: F401  # eager one-time init; see comment above
+
+    # Same class of problem, different torch module. transformers' lazy loader
+    # imports `torch._inductor.test_operators` while resolving classes such as
+    # AutoModelForSequenceClassification / GenerationMixin (exercised by the
+    # cross-encoder / reranker tests). That module registers an `_inductor_test`
+    # TORCH_LIBRARY namespace at module-body level, and under pytest-xdist its
+    # body can execute twice, raising "Only a single TORCH_LIBRARY can be used
+    # to register the namespace _inductor_test". The failure surfaces on
+    # whichever shard runs the reranker tests, masked by transformers as a
+    # misleading "sentence-transformers is required for LocalSTEmbeddings"
+    # ImportError. Seed it once here so the later lazy import is a sys.modules
+    # cache hit and the body never re-executes.
+    import torch._inductor.test_operators  # noqa: F401  # see comment above
 except ImportError:
     pass