Stop importing exir when loading the runtime from Python (pytorch#23301)

Loading the runtime from Python imports the whole exporter.
`portable_lib.py` imports `executorch.exir._warnings` only to get
`ExperimentalWarning` for one warning. Importing any `executorch.exir`
submodule runs `executorch/exir/__init__.py`, which pulls in
`torch._dynamo`, Inductor, Triton and torchao. A process that only loads
and runs `.pte` files pays for all of that.

On a Jetson AGX Thor, in a fresh process after `import torch`, importing
the runtime that way adds about 1,900 modules, over 250 MB of resident
memory and about 3 seconds. Jetson boards share their memory between the
CPU and the GPU, so on a small one like the 8 GB Orin Nano those 250 MB
come out of the model's budget.

This change raises the same message as a `DeprecationWarning`.
`ExperimentalWarning` derives from `DeprecationWarning`, so Python's
default filters treat it the same way: hidden by default, shown with `-W
default` or under pytest. Only the class name in the printed warning
changes.

It also drops the now unused `//executorch/exir:_warnings` dependency
from the `portable_lib` target, and adds a runtime test that imports
`executorch.runtime` in a fresh interpreter and checks that neither
`executorch.exir` nor `torch._dynamo` got loaded.

Measured on the Thor with a model that runs one TensorRT engine through
ExecuTorch, three fresh processes each:

| | before | after |
| --- | --- | --- |
| time to import the runtime and load the program | 3.3 s | 0.24 s |
| resident memory added | 758 MB | 522 MB |
| per-call latency | unchanged | unchanged |

Test plan: the new test fails before this change and passes after it.
The runtime and pybindings tests pass. Checked that the warning still
prints under `-W default` and stays hidden with the default filters.

Author: shoumikhin

extension/pybindings/BUCK

@@ -68,7 +68,6 @@ fbcode_target(_kind = runtime.python_library,
     visibility = ["PUBLIC"],
     deps = [
         ":_C",
-        "//executorch/exir:_warnings",
     ],
 )
 

extension/pybindings/portable_lib.py

@@ -18,11 +18,11 @@
 import sys
 import warnings as _warnings
 
-import executorch.exir._warnings as _exir_warnings
-
+# A DeprecationWarning, like exir's ExperimentalWarning, without importing exir.
+# Importing it would load the exporter into a process that only runs programs.
 _warnings.warn(
     "This API is experimental and subject to change without notice.",
-    _exir_warnings.ExperimentalWarning,
+    DeprecationWarning,
 )
 
 # When installed as a pip wheel, we must import `torch` before trying to import
@@ -107,5 +107,4 @@
 # Clean up so that `dir(portable_lib)` is the same as `dir(_C)`
 # (apart from some __dunder__ names).
 del _torch
-del _exir_warnings
 del _warnings

runtime/test/test_runtime.py

@@ -5,6 +5,8 @@
 # LICENSE file in the root directory of this source tree.
 
 import io
+import subprocess
+import sys
 import tempfile
 import unittest
 from pathlib import Path
@@ -20,6 +22,17 @@
 
 
 class RuntimeTest(unittest.TestCase):
+    def test_runtime_import_does_not_load_the_exporter(self):
+        # A fresh interpreter, since this process has already imported exir.
+        probe = (
+            "import sys; import executorch.runtime; "
+            "print(sorted(m for m in ('executorch.exir', 'torch._dynamo') if m in sys.modules))"
+        )
+        result = subprocess.run(
+            [sys.executable, "-c", probe], capture_output=True, text=True, check=True
+        )
+        self.assertEqual(result.stdout.strip(), "[]")
+
     def test_smoke(self):
         ep, inputs = create_program(ModuleAdd())
         runtime = Runtime.get()