[tknopp]

Removes the pruning of the pass plans: the zero-input and needed-output masks that plan_pass took as
nonzero and needed. The users of the passes prune by themselves (NFFT.jl prunes by boxes of lines and
calls plan_pass(v, d; sign, threads) only), and each mask created its own codelets, so the masks cost
a type parameter, a codelet family and their tests without paying off (the gain measured for r2r was
0.94–0.99 and for in-row pruning 1–5 % of a whole transform).

Changes

• plan_pass: the keywords nonzero and needed are gone. Breaking for code that passes them.
• Removed: Pruning, NOPRUNE, pruning, PRUNE_MIN_ZERO, the mask parameters of the codelet and the
  stage kernels, prune_statements!, the Pr parameter and the prune field of PassPlan, and their tests.
• The codegen no longer represents zero inputs as nothing. The generated codelets are identical to
  those before for every radix from 2 to 31 and both signs (compared statement by statement).
• Documentation and README follow.
• The quick test selection takes about 5 minutes with two Julia threads on a 256-thread machine.

Checks

• Full quick selection (AMD EPYC 7702, two Julia threads): 2683 pass, 10 broken as before, 0 fail.
• Direct DFT check of the quick selection (688 transforms, all radices, both signs): passes.
• Timing against main on the M1, one thread, interleaved, minimum of two runs, 20 cases: 0.97–1.06.

🤖 Generated with Claude Code