optimized/grid_sampler_2d: rename fp16_hw buck lib to avoid op_ prefix

Buck's op_registration_util._enforce_deps rejects any dep starting with
`:op_` on the theory that op_targets should not depend on other
op_targets. The previously-named `:op_grid_sampler_2d_fp16_hw` tripped
that check when op_grid_sampler_2d (which is an op_target) declared it
as a dep.

Rename the internal helper library to `grid_sampler_2d_fp16_hw_impl`,
matching the existing `add_sub_impl` / `binary_ops` naming for
op-specific implementation helpers. No change to file contents, CMake,
or the C++ dispatch — only the Buck target name and the corresponding
dep reference.

Author: jgibson2

kernels/optimized/cpu/targets.bzl

@@ -80,9 +80,11 @@ def define_common_targets():
     # must be a separate translation unit — op_grid_sampler_2d.cpp (the
     # runtime dispatcher) remains on plain ARMv8 and only calls into this
     # after cpuinfo_has_arm_neon_fp16() reports true. Scoped compile flag
-    # stays local to this library.
+    # stays local to this library. Named without the "op_" prefix so the
+    # op_registration_util dependency check (which forbids op_target ->
+    # op_target edges) still lets op_grid_sampler_2d depend on it.
     runtime.cxx_library(
-        name = "op_grid_sampler_2d_fp16_hw",
+        name = "grid_sampler_2d_fp16_hw_impl",
         srcs = ["op_grid_sampler_2d_fp16_hw.cpp"],
         visibility = ["PUBLIC"],
         compiler_flags = select({

shim_et/xplat/executorch/kernels/optimized/op_registration_util.bzl

@@ -224,8 +224,10 @@ OPTIMIZED_ATEN_OPS = (
             # Hardware fp16 path lives in a separate translation unit so the
             # ARMv8.2-a+fp16 compile flag can be scoped locally. A runtime
             # cpuinfo_has_arm_neon_fp16() check in op_grid_sampler_2d.cpp
-            # picks between it and the software-convert fp16 path.
-            ":op_grid_sampler_2d_fp16_hw",
+            # picks between it and the software-convert fp16 path. Named
+            # without the "op_" prefix so _enforce_deps doesn't reject it
+            # as an op_target-to-op_target edge.
+            ":grid_sampler_2d_fp16_hw_impl",
             "fbsource//third-party/cpuinfo:cpuinfo",
             "//executorch/runtime/core/portable_type/c10/c10:aten_headers_for_executorch",
         ],