[fft] Exploit bivariate padding structure in transforms

[devin-ai-integration]

Summary

The Kinoshita–Li buffers are bivariate packings (y = x^C) whose stride-C rows are only half-occupied, i.e. zero wherever (j mod C) >= C/2 — a 2D convolution zero-padded in x. This prunes the FFTs on both sides of that structure, cutting a constant factor off the leading n log² n transform work (the generalization of the earlier evens-upsample trick from b=0 to arbitrary strides).

New engine primitives (documented in the engine contract, required by the concept):

• E::transform_padded(in, n, C) — transform of an input zero on the dead columns. In fft_core::forward_padded, the cross-row DIF stages (k >= C) only touch the live columns (half work), and the stage pairing (j, j+C/2) reads known zeros so it degenerates to copy+twiddle; stages below run in full. Output is the ordinary full transform, so it composes with everything else.
• E::finish_padded(p, out, C, op) — finish when only the live outputs are wanted; fft_core::inverse_padded is the exact mirror (output-pruned inverse), and the dead positions of out are left unspecified. Valid for any product spectrum, since it's output pruning, not an input-support assumption.

Implemented for ntt, real (packing pairs adjacent x's, so the row structure survives at stride C/2 in the packed buffer; C == 2 falls back to the dense path), split, crt, and forwarded through componentwise (matrix/trunc/stable get it for free, with the non-assign op composed through an element buffer as in finish).

Call sites:

// packed_bivariate::advance — Q's x-truncation zeroes the top half of each row
auto tq = E::transform_padded(c, B, 2 << (L - l));
E::finish_padded(E::graeffe(tq, B), c.first(B/2), 1 << (L - l));

// ps_compose pushdown — P's parity mask has the same shape, and the next
// level's mask re-zeroes exactly the positions finish_padded left unspecified
auto tp = E::upsample(E::transform_padded(P.first(B/2), B/2, 1 << (L - l)), B, false);
if (l > 0) E::finish_padded(E::mul(tneg[l], tp, B), P, 2 << (L - l));
else       E::finish(E::mul(tneg[l], tp, B), P);

Tests: all-engine transform_padded/finish_padded equivalence against the plain transform/finish (including the add_op path), plus the same for the matrix/trunc/stable wrapper engines; the existing compose tests cover the series call sites.

Stacked on #87 (uses E::graeffe); base is that branch and should retarget to master once it merges.

Link to Devin session: https://app.devin.ai/sessions/7d41aa62f31e4620807608a6e6afcc94
Requested by: @ecnerwala

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[github-actions]

GCC Code Coverage Report

📂 Overall coverage

Metric	Coverage
Lines	🟡 8763/10590 (82.7%)
Functions	🟢 1211/1318 (91.9%)
Branches	🟡 6933/8821 (78.6%)