Define the PyTorch and InfiniRT compatibility matrix

[voltjia]

Problem

The extension directly links PyTorch C++ libraries and an unversioned
libinfinirt.so, while package metadata currently permits any torch and does
not describe a tested InfiniRT range.

Scope

• Define the initially supported Python, PyTorch, compiler ABI, and InfiniRT
  combinations from measured CI results.
• Add a stable/nightly compatibility matrix after the baseline CPU workflow.
• Encode defensible Python/PyTorch metadata constraints.
• Add a runtime diagnostic for incompatible InfiniRT where the upstream API
  permits it.
• Keep manylinux wheel production and runtime bundling policy out of scope.

Acceptance criteria

• Every declared combination is exercised in CI.
• Unsupported combinations fail installation or import with a useful message.
• README and package metadata state the same compatibility policy.

Depends on #4.

[voltjia]

Compatibility work is now split across focused changes. #24 measured the initial compatibility matrix, #27 declared the supported Python and PyTorch versions, #28 added the PyTorch major.minor import guard, and #31 adds a CXX11 ABI guard before backend registration or native loading.

Issue #12 should remain open after #31. The pinned InfiniRT revision does not expose a runtime version API, so a truthful runtime diagnostic still needs an upstream capability or another defensible mechanism. Nightly compatibility also remains outside the currently declared stable matrix.