docs: correct contributor prerequisites and checkpoint scope

[LauraGPT]

Summary

Correct newcomer-facing problems in CONTRIBUTING.md:

• Replace the obsolete Python 3.8 prerequisite with a Python 3.12 recommendation and explain the current dependency floor.
• Use python -m pip for the selected environment and include the missing AutoModel import in the CPU example.
• Install the optional fastapi[standard] CLI dependencies before the documented fastapi run command.
• Describe the released SenseVoiceSmall checkpoint's five ASR languages, keeping the broader 50+ language research scope separate, as the README already does.

User impact

Contributors should not start from a Python version that cannot resolve the repository requirements, encounter an undefined AutoModel when copying the CPU snippet, or mistake research-family language coverage for the released checkpoint's supported languages.

The separate cold install also reproduced fastapi --help exiting with code 1 because the base FastAPI package does not install its optional CLI. The server instructions now include the prerequisite reported by FastAPI itself.

Model, API, and runtime impact

• No model behavior changes.
• One documentation file only. No dependency bounds, model code, API, Docker, workflow, license, or release changes.

Validation

• On Linux x86-64, created separate clean Python 3.8.20 and 3.12.3 environments and ran uv pip install --dry-run --python <environment>/bin/python --default-index https://pypi.org/simple -r requirements.txt against unchanged repository requirements.
• Python 3.8: dependency resolution fails because the required Torch versions have no compatible cp38 wheel.
• Python 3.12: dependency resolution succeeds with 126 packages. This first check was resolution-only; the separate cold-install checks below were subsequently completed.
• In a separate clean Python 3.12 environment on Linux x86-64, python -m pip install --index-url https://pypi.org/simple -r requirements.txt completed against unchanged requirements (126 packages, 196 seconds).
• python -m pip check: no broken requirements. The guide's python -c "from funasr import AutoModel; print('Installation successful')" completed successfully.
• Imported Torch 2.14.0+cu130 and torchaudio 2.11.0+cu130 together and ran a small CPU tensor operation. This is an import/CPU smoke check, not model inference or GPU compatibility acceptance.
• Reproduced fastapi --help failing with the missing fastapi[standard] prerequisite in the clean requirements-only environment; installing that extra completed successfully (103 seconds).
• Limitations: the subsequent CLI help check timed out without output, as did an independent audio-load probe. Neither is counted as a pass or evidence of a repository defect. Full CLI/service and audio-load acceptance remain unverified; the guide now includes the prerequisite requested by the installed FastAPI package.
• Checked the published Torch 2.12.1 metadata, which declares Requires-Python: >=3.10.
• Checked the README's released-checkpoint scope and SenseVoiceSmall.lid_dict in model.py against the documented language list.
• On the initial documentation commit a73e0f6, python -m pytest -q tests/test_whats_new.py tests/test_canonical_qwenaudio_links.py tests/test_github_templates.py tests/test_funasr_requirement.py: 11 passed.
• On a73e0f6, python3 -m unittest tests.test_container_contract tests.test_device_env: 14 passed.
• On a73e0f6, parsed the Markdown with markdown-it-py, parsed/compiled the Python fence, checked its from funasr import AutoModel, and checked removal of the obsolete prerequisite and checkpoint claim.
• The documentation-test rerun after adding the CLI prerequisite timed out without output; later checks in that sequence were not reached. The earlier test counts are not claimed for the new commit. The follow-up was self-reviewed and checked against the exact intended documentation-only replacement.
• git diff --check.
• Hosted run 35691965301 completed successfully on follow-up head db9d5ec6ff562dcd6062af38f7191d52e18802b4. Its execution log confirms 14 container/device tests and 6 model-timestamp tests passed; Docker context checks and the Linux AMD64 image build also passed. The pull-request workflow did not publish an image. This CI scope does not include the timed-out local documentation-test rerun or CLI/audio-load acceptance checks above.

Notes for reviewers

Self-reviewed, signed and DCO-signed-off commit. A full dependency installation and import smoke checks were subsequently completed as listed above; no model download, CPU/GPU model inference, service acceptance test, or desktop/browser execution was performed. Other Python versions and platforms are not claimed as a tested compatibility matrix. The Orca release-link update in #355 is separate; this branch starts from main and changes only the contribution guide. No user issue is closed by this PR.