chore(annotation): defer _layer_metadata and AnnotationMetadata to follow-up PR

Address PR review comments pytorch#4147:
- Comment 4: Remove AnnotationMetadata / attach_annotation_metadata /
  get_annotation_metadata from _specs.py and __init__.py — unused in this PR
  (designed for @tta.export_as which is deferred).
- Comment 5: Remove _layer_metadata.py and its set_tta_layer_metadata call
  in lower_custom_plugin_descriptor — diagnostic-only (TRT engine inspector),
  non-fatal by design, out of scope for this PR.

Both modules are preserved in the backup branch and will be reintroduced in
a higher-level diagnostics / export_as PR.

Author: Bowen Fu

py/torch_tensorrt/annotation/__init__.py

@@ -30,7 +30,6 @@
 from ._errors import TTADiagnosticError
 
 from ._specs import (
-    AnnotationMetadata,
     KernelImplSpec,
     CuTeDSLSpec,
     CuTileSpec,
@@ -47,7 +46,6 @@
     # Error types
     "TTADiagnosticError",
     # Descriptor types
-    "AnnotationMetadata",
     "KernelImplSpec",
     "TritonSpec",
     "CuTileSpec",

py/torch_tensorrt/annotation/_custom_plugin/_descriptor.py

@@ -44,7 +44,6 @@
     make_td_from_meta_using_template,
     make_td_from_meta_using_template_symbolic,
 )
-from .._layer_metadata import set_tta_layer_metadata
 from .._specs import CuTeDSLSpec, CuTileSpec, TritonSpec
 from ._symbolic import SymbolicTensor, TensorRole
 
@@ -1250,26 +1249,6 @@ class docstring).  It:
         # float(np.array([v])) fails in NumPy 1.25+).
         plugin_layer = ctx.net.add_plugin(plugin_fn(*trt_inputs), aot=True)
         plugin_layer.name = name
-        _first_spec = descriptor.specs[0] if descriptor.specs else None
-        if isinstance(_first_spec, TritonSpec):
-            _backend = "triton"
-        elif isinstance(_first_spec, CuTileSpec):
-            _backend = "cutile"
-        elif isinstance(_first_spec, CuTeDSLSpec):
-            _backend = "cutedsl"
-        else:
-            _backend = "custom_plugin"
-        _fn_specs = []
-        for _s in descriptor.specs:
-            _fn = getattr(getattr(_s, "launch_fn", None), "__name__", None)
-            if _fn:
-                _cfgs = _s.configs if getattr(_s, "configs", None) else [{}]
-                for _cfg in _cfgs:
-                    _fn_specs.append((_fn, _cfg))
-        set_tta_layer_metadata(plugin_layer, _backend, plugin_name, name,
-                               attrs=descriptor.attrs or None,
-                               fn_specs=_fn_specs or None)
-
         if plugin_layer.num_outputs == 1:
             return plugin_layer.get_output(0)
         return tuple(