feat(annotation): add TTA annotation layer (export_as, lower_as, custom_plugin)

Adds torch_tensorrt.annotation (aliased as tta) — a zero-overhead annotation
layer that lets users tag regions of a PyTorch model for custom TensorRT
lowering, without modifying core torch_tensorrt internals beyond a small set
of generic extension hooks.

Core hooks added to torch_tensorrt (outside annotation/):
- _compiler.py: generic extension hook registries (register_compile_pass,
  register_preserved_ep_attr, register_export_context, register_post_trace_hook)
  + EP attribute preservation across run_decompositions() + compile pass loop
- _compile.py: post-trace hook loop
- _settings.py: profiling_verbosity field in CompilationSettings
- _tracer.py: ExitStack wrapping for registered export context factories
- _ConversionContext.py: current_node field for converter access
- _TRTInterpreter.py: PREFER_AOT_PYTHON_PLUGINS flag, profiling_verbosity
  routing, current_node tracking, layer metadata stamping

Annotation module (py/torch_tensorrt/annotation/):
- export_as: context manager to tag regions during tracing
- lower_as: context manager to specify custom TRT lowering (builtin/plugin/kernel)
- custom_plugin: Triton/CuTile/CuTeDSL AOT QDP plugin descriptors
- IR layer: region discovery, boundary validation, region views
- Full test suite (599 unit/integration tests + 25 pre-Blackwell tests)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: Bowen Fu

py/torch_tensorrt/_compile.py

@@ -320,6 +320,12 @@ def _fx_input_interface(
                 kwarg_inputs=torchtrt_kwarg_inputs,
                 **kwargs,
             )
+        # Run post-trace hooks.
+        from torch_tensorrt.dynamo._compiler import _post_trace_hooks
+        for _hook in _post_trace_hooks:
+            _result = _hook(exp_program, torchtrt_arg_inputs)
+            if _result is not None:
+                exp_program = _result
         trt_graph_module = dynamo_compile(
             exp_program,
             arg_inputs=torchtrt_arg_inputs,