Prototype: TensorRT execution provider (linux, CUDA 13.0)

[diegoferigo-rai]

Draft / prototype — not meant to be merged. TensorRT packages come from a
local channel, so CI here cannot build it. Scoped to linux; posted as a
starting point to adapt. Relates to #199.

Approach

Optional TensorRT execution provider on the CUDA 13.0 linux builds
(linux-64 and linux-aarch64), following the pytorch-feedstock style of
encoding the accelerator in the build string. A dedicated tensorrt variant
axis (linux) gates it; the _tensorrt marker only appears on the TRT build:

Variant	Build string
CUDA 13.0 + TRT	cuda130_tensorrt_py310_h…_200
CUDA 13.0 (no TRT)	cuda130_py310_h…_200
CUDA 12.9	cuda129_py310_h…_200
CPU	cpu_py310_h…_0

The inert tensorrt: true combinations (CPU / CUDA 12.9) are skipped, so no
duplicate packages. The CI matrix is restricted to linux for the prototype.

Changes

• recipe/recipe.yaml: tensorrt_enabled context gated to CUDA 13.0 via the
  tensorrt axis; _tensorrt appended to the build string only on the TRT build;
  TensorRT host deps libnvinfer-devel + libnvonnxparser-devel under
  if: tensorrt_enabled (run_exports pin the runtime libs); package_contents
lib check for onnxruntime_providers_tensorrt; skip not linux.
• recipe/build-cpp.nu: cmake flags onnxruntime_USE_TENSORRT=ON,
  onnxruntime_USE_TENSORRT_BUILTIN_PARSER=ON, onnxruntime_TENSORRT_HOME=$PREFIX.
  Builtin parser -> shared libnvonnxparser, nothing static.
• recipe/conda_build_config.yaml: tensorrt axis (linux) + local
  channel_sources for the locally-built TensorRT packages.

To adapt

• Replace the local channel_sources
  (file:///…/staged-recipes/build_artifacts,conda-forge) with the real TensorRT
  packages once they land on conda-forge.
• The local channel only ships linux-64 artifacts; the CUDA 13.0
  linux-aarch64 TRT build won't solve until aarch64 packages are built.
• Restore the non-linux platforms and decide how far to extend TensorRT.
• Confirm the builtin (dynamic libnvonnxparser) vs. static parser choice.

[conda-forge-admin]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/recipe.yaml) and found it was in an excellent condition.

I do have some suggestions for making it better though...

For recipe/recipe.yaml:

• ℹ️ output 0 output overrides versions pinned in the feedstock:
  ['- In section host: python 3.13.*']
  Requirement spec should not list version specifiers to respect conda-forge-pinning. If you need to force another version, please override the pin via conda_build_config.yaml.
• ℹ️ License is not an SPDX identifier (or a custom LicenseRef) nor an SPDX license expression.

Documentation on acceptable licenses can be found conda-forge.org > Docs > Maintainer Documentation > Contributing packages > SPDX Identifiers and Expressions.

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/31599157505. Examine the logs at this URL for more detail.}