fix(ci): unbreak rerankers (torch bump) and vllm-omni on aarch64

[localai-bot]

Summary

Two independent CI breakages on master, bundled here because each fix is a one-liner.

rerankers backend tests fail (cpu / cublas12)

requirements-cpu.txt and requirements-cublas12.txt pin torch==2.4.1 but leave transformers unpinned. The latest transformers (5.x) registers a custom op in transformers/integrations/moe.py:

```python
torch.library.custom_op("transformers::grouped_mm_fallback", _grouped_mm_fallback, mutates_args=())
```

`_grouped_mm_fallback`'s signature uses string-typed annotations ('torch.Tensor'). torch 2.4.1's infer_schema does not understand them and raises ValueError: Parameter input has unsupported type torch.Tensor. That import error prevents the gRPC server from starting, so all 5 rerankers tests fail with Connection refused against 127.0.0.1:50051.

Fix: bump torch==2.4.1 → torch==2.7.1. Matches the pin used by the transformers backend in this repo and is the most common torch version across our Python backends. Leaves transformers unpinned so we keep tracking upstream.

vllm-omni build fails on aarch64

vllm-omni's setup.py resolves dependencies dynamically and loads requirements/cuda.txt on cuda hosts, which pins fa3-fwd==0.0.3. fa3-fwd ships only manylinux_2_24_x86_64 wheels and has no source distribution, so on aarch64 (e.g. l4t13 / SBSA cu130) uv fails with:

```
Because fa3-fwd==0.0.3 has no wheels with a matching platform tag (e.g.,
`manylinux_2_39_aarch64`) and vllm-omni==... depends on fa3-fwd==0.0.3,
we can conclude that vllm-omni cannot be used.
```

Fix: in install.sh, after cloning vllm-omni, strip fa3-fwd from requirements/cuda.txt when building on aarch64. This is safe — fa3-fwd is a soft runtime dep:

• vllm_omni/diffusion/attention/backends/utils/fa.py wraps from fa3_fwd_interface import ... in try/except ImportError and falls back through FA3 source build → FA2 (flash_attn) → vLLM's wrapper.
• vllm_omni/diffusion/attention/backends/flash_attn.py only raises if every FA backend is missing, with a message offering SDPA as a final fallback.

Test plan

• CI green on tests-apple / tests-linux for the rerankers backend
• aarch64 build (l4t13 / SBSA) reaches the pip install -e . step for vllm-omni without resolver failure

Assisted-by: Claude:claude-opus-4-7-1m [Claude Code]