fix: stop calling TensorRT 10.15 aliasing APIs unconditionally

[shoumikhin]

The problem

Three things Torch-TensorRT calls without checking were added in TensorRT 10.15:

• ICudaEngine::getAliasedInputTensor (C++)
• ICudaEngine.get_aliased_input_tensor (Python)
• the KV-cache update layer, with its KVCacheMode enum

Building against an older TensorRT fails outright:

core/runtime/TRTEngine.cpp: error: 'class nvinfer1::ICudaEngine'
    has no member named 'getAliasedInputTensor'

Once that is past, compiling any model fails, not only one with a cache:

File "py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py", line 519, in run
    aliased_in = cuda_engine.get_aliased_input_tensor(out_name)
AttributeError: 'ICudaEngine' object has no attribute 'get_aliased_input_tensor'

That loop runs after every engine build, over every output, so an ordinary model with no
aliasing at all hits it. The same call on the Python runtime fails the same way at engine
load.

The fix

Reconciliation compares the build-time alias map against what an engine reports about
itself. A build without the API cannot report anything, so there is nothing to compare
and the build-time map already stands alone. Skip only that loop, and keep everything
after it, because user-declared aliases exist independently of the TensorRT API and their
binding names still need recomputing.

Writing through an alias is a different matter, because it cannot be emulated. Where the
code promises a write-through, it now refuses rather than quietly dropping it. Two paths
promise one:

• A cache passed in as a network input and written in place. Falling back to the
  functional scatter computed the correct return value and never touched the caller's
  tensor. That is worse than an error, because a test that only checks numbers passes
  while the cache silently stops updating.
• A mutated module buffer. Lifting removes the writeback copy, on the assumption the
  engine writes through the aliased input.

Both now raise, naming the input or the buffer and the version required:

Writing in place to the network input 'cache' needs the TensorRT KV-cache update
layer, which was added in TensorRT 10.15. This build does not have it, and without
it the write would be computed and then discarded. Use TensorRT 10.15 or newer, or
return the updated tensor instead of writing into an input.

One shared capability check backs both sites, so they cannot drift apart.

Scope

This does not make Torch-TensorRT work on TensorRT older than 10.15. Other 10.15-only
symbols are still referenced unconditionally, including trt.IAttention in an annotation
that is evaluated at import time, so importing the package on 10.13 fails before any of
this runs. What this change does is take these three APIs off that list.

Testing

ordinary model, aliasing API removed        compiles, matches eager
module buffer write, KV layer removed       raises, names the buffer
network input write, KV layer removed       raises, names the input
unmodified TensorRT 11.1.0.106              unchanged, matches eager
tests/py/dynamo/runtime/test_aliased_io.py  13 passed
black --check on all five changed files     clean

The three added tests fail without this change and pass with it. Also confirmed on an
aarch64 device with system TensorRT 10.13.3.9 that all three APIs are absent there, and
that the C++ version check evaluates false on it.

[narendasan]

@shoumikhin What Jetpack version are you using?

[shoumikhin]

JetPack 7.1 (L4T R38.4, CUDA 13.0, aarch64), whose apt channel ships TensorRT 10.13.3.9, where getAliasedInputTensor is absent from the installed headers and trt.KVCacheMode / add_kv_cache_update are missing from the Python module.