Only defer top-level functions

[ilevkivskyi]

This makes deferral logic more robust and more consistent with fine-grained mode. I also:

• Change some terminology, as "top function" is ambiguous: top-level function vs top of stack function.
• Update some docs and type annotations to match actual behavior (e.g. we do not defer lambdas)

See also #18674 for some more motivation.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

steam.py (https://github.com/Gobot1234/steam.py)
- steam/abc.py:667: error: Argument 3 to "_from_proto" of "Review" has incompatible type "PartialUser"; expected "User | ClientUser"  [arg-type]

static-frame (https://github.com/static-frame/static-frame)
+ static_frame/core/archive_npy.py:1185: error: Unused "type: ignore" comment  [unused-ignore]

spark (https://github.com/apache/spark)
+ python/pyspark/util.py:434: error: Unused "type: ignore" comment  [unused-ignore]
+ python/pyspark/util.py:458: error: Unused "type: ignore[misc, operator]" comment  [unused-ignore]
+ python/pyspark/pandas/frame.py:6413: error: Argument "value" to "replace" of "Series" has incompatible type "int | float | str | list[Any] | tuple[Any, ...] | dict[Any, Any]"; expected "list[Any] | tuple[Any, ...]"  [arg-type]
+ python/pyspark/pandas/namespace.py:2288: error: Unused "type: ignore" comment  [unused-ignore]
+ python/pyspark/pandas/namespace.py:2291: error: Unused "type: ignore" comment  [unused-ignore]

[ilevkivskyi]

The only new error looks correct, as now we apply narrowing from binder (Any -> something more precise).