Phase 3 — plan reuse, streaming large FFTs, NEFF cache

[scttfrdmn]

Roadmap phase tracker

This issue tracks trnfft's work on Phase 3 of the trnsci roadmap.

See the suite-level roadmap
for the full phase matrix and cross-project dependencies. A reader-friendly
version lives at trnsci.dev/roadmap/.

What this phase means for trnfft:
Make the NKI path meaningfully faster than the PyTorch fallback.

Done means:

• FFTW-style plan cache validated end-to-end — repeated calls at the same size skip planning.
• Streaming large-N FFT that doesn't materialize intermediate arrays in HBM.
• Published benchmarks at trnsci.dev/trnfft/benchmarks/ with CPU (torch), vintage-matched GPU (A10G / H100), and Trainium numbers.
• NEFF compile-cache reuse verified — first call per shape compiles, subsequent calls hit cache.

Coordination

• Label: phase-3-perf — matches the same label in every sub-project.
• Cross-project dependencies, if any, are called out in the "Done means"
  list above. Link any PRs / child issues here for tracking.

Close this issue when all "Done means" items are satisfied.