build: support Python 3.13 conversion dependencies

[darsh7807]

Summary

Fix installation on Python 3.13 for the conversion dependency set used by BitNet's top-level requirements.txt.

Root cause

BitNet's top-level requirements included vendored llama.cpp requirement files that pin torch~=2.2.1. That torch line does not provide cp313 wheels, so uv pip install -r requirements.txt fails on Python 3.13 with an unsatisfiable dependency resolution error.

Changes

• Route top-level conversion requirement includes through BitNet-local wrapper files instead of referencing the vendored llama.cpp files directly
• Keep torch~=2.2.1 for Python <3.13
• Use torch>=2.5.0,<3.0.0 for Python >=3.13
• Apply the same logic to the LoRA conversion path so top-level installs remain consistent and do not reintroduce the old torch pin transitively

Why this approach

The torch incompatibility exists in vendored submodule requirement files, but BitNet consumes those files through its own top-level requirements.txt. Using BitNet-local wrappers fixes installation behavior without modifying the submodule directly.

Verification

Before

On clean main:

git submodule update --init --recursive
uv venv --python 3.13 --seed .venv-py313
uv pip install --python .venv-py313/bin/python -r requirements.txt

Result:

• fails on Python 3.13
• resolver reports that torch~=2.2.1 has no matching cp313 wheels

After

On this branch:

git submodule update --init --recursive
uv venv --python 3.13 --seed .venv-py313
uv pip install --python .venv-py313/bin/python -r requirements.txt

Result:

• succeeds on Python 3.13
• resolves torch==2.10.0

Regression check:

uv venv --python 3.12 --seed .venv-py312
uv pip install --python .venv-py312/bin/python -r requirements.txt

Result:

• still succeeds on Python 3.12
• resolves torch==2.2.2

Closes #413

[darsh7807]

@microsoft-github-policy-service agree