[tknopp]

Based on #24 (removal of the pruning); the diff here is the run-time pass plans only.

NFFT.jl stores its passes in containers of one plan type. The plans of plan_pass, plan_r2c_pass and
plan_c2r_pass have a type that depends on the length (the kernel types of the length), so such a
container cannot be compiled with juliac --trim. This pull request adds run-time versions whose type
depends on the element type and the number of dimensions only:

• plan_pass_runtime(v, d), plan_r2c_pass_runtime(v, d), plan_c2r_pass_runtime(v, d): the same
  transforms as the specialised functions, with run-time kernels only (RChain; Bluestein and composite over
  RChain; four-step over RChain). The plan type is PassPlan{T,N,RPassKern{T}} resp.
  RealPassPlan{T,N,RealLine{T,RKern{T}}} for every length.
• The specialised functions are unchanged in their kernels and their results.
• The line offsets of a pass are built by doubling, which makes planning faster (see below).
• The trimmable plans are described in the how-to on compiling ahead of time, and the decision is in the
  design decisions page.

Why separate functions and not a keyword

The first version had a keyword runtime = true. The trim verifier checks the code of both branches, and the
specialised branch has calls whose types depend on the length, so the verifier refused it. A constant
keyword is not folded before verification, which Base.@constprop did not change.

Checks

• Trim probe (juliac --trim=safe, the same invocation as libfftl/build.sh): a library with a vector of
  plan_pass_runtime plans and a vector of plan_r2c_pass_runtime plans, both applied with mul!. The
  verifier reports no errors. Called from C, both entry points agree with the direct DFT to 4·10⁻¹⁴ (n = 64
  and n = 100, several columns).
• Quick test selection (AMD EPYC 7702, two Julia threads): 2739 pass, 10 broken, 0 fail. New test set "run-time pass plans" (plan types for
  several lengths, agreement with the specialised plans for smooth, Bluestein, composite and four-step lengths,
  strided views in dimensions 2 and 3, and a direct DFT check).
• Planning (M1, 64 × 64 × 64 views, one thread): the specialised plan takes 4.7 µs for dimension 1 and
  2.8 µs for dimension 2 on main, and 1.8 µs and 0.9 µs with this change. Applying a plan is unchanged within
  the noise of the measurement.

🤖 Generated with Claude Code

[tknopp]

Replaced by a new draft pull request: the branch was rebased onto main after #24 was merged (merge commit), and the history of the old branch was not rewritten. The content is the same plus the reference page for the new functions (fixes the failing documentation build).