fix: stabilize interleaved FFT CPU execution

[shinaoka]

Summary

• add final cube synchronization after interleaved CFFT/RFFT/IRFFT shared-memory stores
• cap the normal CPU shared-memory path test while preserving the real device-limit test under heavy
• use checked arithmetic when constructing unsupported FFT lengths so CPU backends with system-RAM limits skip impossible cases

Root cause

CubeCL CPU can reuse shared memory when workers advance to the next cube. Without a final barrier, multi-window interleaved transforms were nondeterministically corrupted. The CPU backend also reports system RAM as shared-memory capacity, making the ordinary boundary test attempt an impractically large FFT.

Verification

• RUSTUP_TOOLCHAIN=1.97.1 cargo xtask test --ci
• cargo +1.97.1 test -p cubek-fft --features cubecl/wgpu
• cargo +1.97.1 test -p cubek-fft --features cubecl/wgpu,heavy
• cargo +1.97.1 fmt --all -- --check
• cargo +1.97.1 clippy --no-deps -- --deny warnings

Refs #6