"Build without cuda runtime library" approach 1

[nimdrak]

Fixes: #643

What this solves

• building. imports onnx_tensorrt/init.py just to read version.
• Because__init__ imports backend, this transitively pulls in GPU-dependent code.
• As a result, the build appears to “require” GPU access even though it only needs the version string.

Why this matters

• We shouldn’t need a GPU to build the package (especially in containerized CI).
• Removing the implicit GPU requirement lets us build anywhere, making it easier and more efficient to produce and distribute container images.

How to test

• Build the project without a GPU.
• Use a plain runc runtime (not nvidia-container-runtime) as demonstrated in the linked issue comment.
• I also verified that inference still works correctly in a GPU environment at my workplace.

Approach

• There are a few ways to obtain the version without importing heavy modules at build time:
• My recommendation is 1, this PR.

1. Read the version file directly (no import) ← This PR implements this

• Pros: Very simple; minimal change to the current layout.
• Cons: Feels a bit odd to parse Python source; but it’s a common, practical pattern.

2. Use pyproject.toml (PEP 517/518) ← I also implemented here #1036, just for reference.

• Pros: Standards-based; clean separation of build metadata; no import side effects during build.
• Cons: Touches more files than (1).

3. importlib.metadata.version or pkg_resources

• Not suitable here because they assume the distribution is already installed, which isn’t true at build time.

# Not applicable during build:
from importlib.metadata import version
VERSION_NUM = version("InstalledPackageName")

# or
import pkg_resources
version = pkg_resources.require("MyProject")[0].version