fix(runtime): back a python engine runtime cache with the python handle

[shoumikhin]

What is broken

Two tests are red on main, RTX only:

• runtime/test_aliased_io.py::TestPythonRuntimeAliasedIO::test_in_place_write_through
• runtime/test_aliased_io.py::TestPythonRuntimeAliasedIO::test_streaming_state_accumulates

TypeError: set_runtime_cache(): incompatible function arguments.
  1. (self: IRuntimeConfig, cache: nvinfer1::IRuntimeCache = None) -> bool
Invoked with: <IRuntimeConfig object at 0x...>, None

Why

RuntimeCache chooses its backing from whether the C++ library is loaded:

if ENABLED_FEATURES.torch_tensorrt_runtime:
    self._handle = torch.classes.tensorrt.RuntimeCacheHandle(path)
else:
    self._handle = _RuntimeCacheHandle(path=path)

A loaded library does not mean a C++ engine is the consumer. These tests build
a python TRTEngine straight from a packed engine tuple, which is a documented
way to use it. Its default settings carry a cache path string, so
TRTRuntimeConfig._apply_settings wraps that path, gets the torchbind handle,
and asks it for a live cache. The torchbind handle materializes its cache on
the C++ side and cannot pass an IRuntimeCache back to python, so
ensure_cache returns None, and None is not a cache.

Fix

Let the caller say which runtime will consume the cache. _apply_settings
only ever runs for a python engine, so it asks for the python backing. Nothing
else changes: every existing caller keeps the old default.

One related detail: IRuntimeConfig does not own the cache attached to it, so
the wrapper now lives on the config instead of being dropped at the end of the
call. Before this change the wrapper was released immediately, which was
harmless only because the attach never happened.

Tested

The RTX path cannot run on this machine. What was checked here, with the C++
library loaded:

• default backing: ensure_cache returns None, which is the exact input the
  error message reports
• python backing: ensure_cache returns the live cache

test_000_runtime_cache.py, test_004_runtime_settings.py and
test_aliased_io.py give the same result before and after the change on
non-RTX TensorRT, 28 passed and 27 skipped.