fix(deps): align torchao constraint in requirements.txt with pyproject.toml (#1298)

requirements.txt declares `torchao` unpinned while pyproject.toml declares
`torchao>=0.16.0,<0.17.0` for non-aarch64. A fresh
`pip install -r requirements.txt` on Windows therefore resolves torchao 0.18.0,
while the same file pins torch==2.7.1+cu128 for win32.

torchao 0.18.x imports `ScalingType` from `torch.nn.functional`, which does not
exist in torch 2.7.1, so loading the DiT checkpoint fails with:

    ImportError: cannot import name 'ScalingType' from 'torch.nn.functional'
    RuntimeError: Failed to load model with attention implementations
    ['sdpa', 'eager']

Installing the range already declared in pyproject.toml (torchao 0.16.0)
restores loading with no other change. Verified in a clean virtual environment
on Python 3.12.10: torchao resolves to 0.16.0, `import torchao.quantization`
succeeds, the released checkpoint loads on CUDA, and seeded generation produces
decoded audio identical to the same seeded request in a separately built
environment.

Mirrors the pyproject.toml markers exactly so the two declarations cannot drift
apart again.

Author: BigBirdReturns

requirements.txt

@@ -34,7 +34,12 @@ python-multipart>=0.0.18
 numba>=0.63.1
 vector-quantize-pytorch>=1.27.15
 torchcodec>=0.9.1; platform_machine != 'aarch64'
-torchao
+# Must match pyproject.toml. Unpinned, this resolves to torchao 0.18.x, which
+# imports torch.nn.functional.ScalingType -- a symbol that does not exist in the
+# torch 2.7.1+cu128 build pinned above for Windows. The model then fails to load
+# with "cannot import name 'ScalingType'". Keep these two declarations aligned.
+torchao>=0.16.0,<0.17.0; platform_machine != 'aarch64'
+torchao; platform_machine == 'aarch64'
 toml
 modelscope