ENH Add CUDA support for Hilbert transforms

[viranovskaya]

Related to #13993 and #13994.

This implements the alternative suggested in the discussion: CUDA support in apply_hilbert instead of a second public job-control parameter in annotate_muscle_zscore.

When n_jobs="cuda" is requested, the analytic signal is computed with CuPy FFT and transferred back to NumPy. If CUDA cannot be initialized, the existing CPU implementation is used with one job. This allows annotate_muscle_zscore(..., n_jobs="cuda") to use the same setting for filtering and the Hilbert transform without exposing an implementation-specific parameter.

The numerical tests compare the FFT implementation with scipy.signal.hilbert for odd and even lengths, zero padding, analytic signals, and envelopes. A regression test covers the full muscle-annotation call.

Local checks:

• 290 passed, 12 skipped in the filter and artifact-detection test modules
• ruff check and format
• ty
• codespell
• rstcheck

This was tested on a machine without CUDA hardware. The CPU fallback and numerical implementation are covered locally, but the actual CuPy device path still needs CUDA-backed verification.