feat(fft): add interleaved C32 FFT APIs to t4a branch

[shinaoka]

Summary

• add public interleaved C32 CFFT plus interleaved F32/C32 RFFT and IRFFT APIs
• preserve split scratch internally while using enum normalization and the existing device-derived small/large threshold
• include padded RFFT/IRFFT launch paths, caller-owned outputs, alias/layout validation, and PR fix: stabilize interleaved FFT CPU execution #8 CPU stabilization barriers
• pin the reviewed host-visible CubeCL revision from feat(wgpu): add guarded host-visible primary memory cubecl#13

Scope

Metal remains F32/C32 only. F64/C64 are intentionally not added to CubeK. Benchmark catalog backport and crate release are deferred; tenferro can use this exact Git revision.

Verification

• cargo test -p cubek-fft --all-features: 2 unit + 77 integration passed
• cargo clippy -p cubek-fft --all-targets --all-features -- -D warnings
• fmt and diff check
• independent re-review: no Critical/Important findings

Closes #6.
Supports tensor4all/tenferro-rs#1394.