Support PyTorch 2.11.0 and PyTorch 2.13.0 with CUDA 13

[avidale]

Motivation

Adds PyTorch 2.13.0 (cpu, cu130) to the support matrix, which enables co-installation with vllm==0.28.0 (pins torch==2.13.0 plus CUDA 13 runtime packages) and transformers~=5.16, removing friction in the downstream projects, such as Omnilingual.

Changes

Three things had to change to build against PyTorch 2.13:

• PyTorch 2.12 and later set CMAKE_CXX_STANDARD 20, so fairseq2n is now compiled as C++20. This is a breaking change for source builds: GCC 11.3 or greater, or Clang 16 or greater, is now required. The CI images already ship gcc-toolset-14, so _build_wheel-linux.yaml selects it for PyTorch 2.12+ and keeps gcc-toolset-11 (GCC 11.2) for older versions.

• Under C++20, fmt::format_string is checked at compile time, so runtime strings can no longer be passed as the format. sndfile.cc and sp_processor.cc now pass libsndfile and SentencePiece error text as an argument instead. This also fixes a latent bug: error text containing braces was previously interpreted as format syntax.

• Also under C++20, the unqualified ssize() call in the memory bindings became ambiguous with std::ssize, because std::byte in the template argument pulls std into the ADL set. It is now qualified.

CUDA 13 dropped Volta, so the default CMAKE_CUDA_ARCHITECTURES is Turing rather than Volta when PyTorch is built against CUDA 13, and the wheel workflow selects architectures per variant.

Finally, the bundled zip third-party project compiles itself with -Werror, and GCC 14's new -Wcalloc-transposed-args makes that fatal, so -Werror is disabled for that target only.

Verified for torch 2.13.0+cu130 / py3.12 / x86_64: native suite 28/28, pytest --device cpu 1155 passed, pytest --device cuda:0 on an H100 1154 passed (the one failure, gemma3n test_batch_independence, is a pre-existing GPU float-tolerance issue: max abs diff 9.5e-07 with atol left at its 1e-8 default). Wheels built in the new manylinux_2_28 cu130 image audit clean at glibc 2.27.

Some additional changes were required to make the Github CI pass; they seem unrelated to the torch/cuda/transformers versions upgrade, and instead, chase the natural evolution of the other dependencies:

• Change the Python version for the mypy check from 3.10 to 3.12, otherwise, the type stubs for numpy>=2.5, scipy, librosa are incompatible with the syntax
• adapt to a refactoring of the wandb dependency: wandb.util.generate_id → from wandb.sdk.lib.runid import generate_id

Does your PR introduce any breaking changes? If yes, please list them:
List of all backwards-incompatible changes.

Check list:

• Was the content of this PR discussed and approved via a GitHub issue? (no need for typos or documentation improvements)
• Did you read the contributor guideline?
• Did you make sure that your PR does only one thing instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (no need for typos, documentation, or minor internal changes)