Remove arithmetic transforms from InSampleUniform registration (facebook#5117)

Summary:
Pull Request resolved: facebook#5117

`InSampleUniformGenerator` selects existing arms for LILO labeling -- it does not generate new points and therefore does not need parameter normalization.

The `Cont_X_trans` transform chain includes `Log`, `Logit`, and `UnitX`, which perform arithmetic whose forward/reverse round-trip introduces ~1e-15 IEEE 754 rounding. This can produce arms with slightly different parameter values and signatures ("ghost arms"), preventing the experiment from reusing existing `Arm` objects. Downstream, LILO source resolution fails because no arm matches by signature.

This diff registers `InSampleUniform` with only the structural/type transforms (`RemoveFixed`, `OrderedChoiceToIntegerRange`, `OneHot`, `IntToFloat`) needed for categorical, ordered-choice, integer, and fixed parameter compatibility. The arithmetic transforms (`Log`, `Logit`, `UnitX`) are removed since they serve no purpose for in-sample selection and introduce precision loss.

Note: for experiments with `digits` set on `RangeParameter`, the auto-added `Cast` transform also rounds parameter values. In properly configured new experiments this is consistent (Cast untransform re-rounds on the reverse path), but historical experiments may have arms stored with un-rounded values from before Cast handled this correctly. This diff addresses the UnitX/Log/Logit source of precision loss; historical data inconsistencies are a separate concern.

Reviewed By: andycylmeta

Differential Revision: D98970704

fbshipit-source-id: e3baeb66156393f81492bd02208bcaac3f7b9aa2

Author: ItsMrLin

ax/adapter/registry.py

@@ -216,10 +216,17 @@ class GeneratorSetup(NamedTuple):
         generator_class=UniformGenerator,
         transforms=Cont_X_trans,
     ),
+    # In-sample generators only select existing arms -- they do not need
+    # arithmetic transforms (Log, Logit, UnitX) whose forward/reverse
+    # round-trip introduces ~1e-15 IEEE 754 rounding.  This rounding
+    # produces "ghost arms" with slightly different parameter values
+    # and signatures, preventing the experiment from reusing existing
+    # Arm objects.  Structural/type transforms are kept for compatibility
+    # with categorical, ordered-choice, integer, and fixed parameters.
     "InSampleUniform": GeneratorSetup(
         adapter_class=RandomAdapter,
         generator_class=InSampleUniformGenerator,
-        transforms=Cont_X_trans,
+        transforms=[RemoveFixed, OrderedChoiceToIntegerRange, OneHot, IntToFloat],
     ),
     "ST_MTGP": GeneratorSetup(
         adapter_class=TorchAdapter,