[sunyi0505]

Summary

• torch.compile is natively supported in torch_npu==2.7.1.post4. Remove unused triton-ascend GeLU and SquaredReLU NPU kernels (gelu_fwd/bwd, sqrelu_fwd/bwd), including Triton JIT kernels, Python wrappers, and TritonAscendBackend dispatch methods.
• Drop @dispatch('modules') from gelu_fwd, gelu_bwd, sqrelu_fwd, and sqrelu_bwd in fla/modules/activations.py; these ops now use the @torch.compile default path on all devices, including NPU.
• Add tests/utils/test_ascend_ub_manager.py with unit tests and NPU integration tests for ascend_ub_manager (UB capacity detection, tiling strategy, block-size helpers, grid-axis chunking).

Motivation

GeLU/SquaredReLU NPU kernels are not used in practice and add maintenance overhead. UB manager logic from #1001 lacked test coverage.

[sunyi0505]

This PR is ready. Could you please review it? @zhiyuan1i

[zheliuyu]

Thanks @ChengQianqian for reporting this issue.