perf(stark): GPU transfer-residency

[jotabulacios]

GPU transfer-residency: keep composition + FRI + DEEP resident on device

Motivation

The GPU prover shipped large intermediates across PCIe every prove: the extended
composition LDE crossed the bus three times (D2H, re-upload for Merkle, re-upload
for DEEP), the FRI layer evals were copied to host for the query phase, and the
DEEP→FRI bridge did D2H + CPU bit-reverse + H2D before the first fold. The GPU
sat idle waiting on these round-trips.

Description

Keep the intermediates on the device; gather only what the openings need.

• Composition: retain the extended-LDE buffer as a typed handle and build the
  Merkle tree, R3 OOD, and R4 openings from it; drop the host D2H when a table's
  DEEP inputs are all resident (others keep the host fallback). Cross-stream
  ordering via a recorded "LDE-ready" event.
• FRI: keep each committed layer's folded evals resident; gather queried
  opening values from device instead of copying full layers to host.
• DEEP→FRI: pass the natural-order codeword to FRI as a resident handle; the
  first fold reads br(2j)/br(2j+1), folding the bit-reverse into the fold and
  removing the D2H + CPU bit-reverse + H2D bridge.

[jotabulacios]

/bench-gpu

[github-actions]

GPU Benchmark (ABBA) — 738f79cf8d vs main (14 pairs)

_{RTX 5090 · Vast.ai datacenter @ $0.8423148148148147/hr · prover/cuda · drift-free A/B/B/A}

=== ABBA paired result  (improvement: - = PR faster) ===
  pairs: 14   mean A (PR): 13.573s   mean B (base): 15.293s

  [parametric] paired-t   mean -11.24%   sd 1.24%   se 0.33%
               95% CI: [-11.96%, -10.52%]   (t df=13 = 2.16)
  [robust]     median -11.52%   Wilcoxon W+=0 W-=105  p(exact)=0.0001  (z=-3.26)

  --- server stability (this run; compare across servers) ---
  run-to-run jitter:    A CV 1.01%   B CV 1.31%        (lower = steadier)
  within-session drift: +2.66% over the run, 1st->2nd half +1.30%
    (jitter -> Tier-1 cached gate floor; drift -> whether the cached baseline can be trusted)

  VERDICT: REAL IMPROVEMENT - PR faster by ~11.24% (t-CI and Wilcoxon agree)

  raw pairs: /tmp/abba_run/pairs.csv

_{- = PR faster. Trust the verdict when paired-t and Wilcoxon agree.}