fix: thread a real per-channel weight stride through the field _jrls drivers (CUDA)

[balbasty]

Summary

Structural mirror of the fastfields-cpu-impl driver fix for the same bug
(fastfields-cpu-impl#53). The field regulariser's _jrls CUDA driver
layer (matvec/diag/relax x absolute/membrane/bending, 9 functions) never
derived a per-channel weight stride at all: the CUGLOB kernels' local
stride_wgt stack copy was only nall long (one short of the
channel-stride entry at index nall), and the CUHOST launchers only
uploaded nall elements to the device -- so there was no wsc to compute
in the first place, and every channel implicitly read wgt[wgt_offset]
(channel 0's weight) via the same base pointer.

Fix, mirroring the RLS (wc=1) drivers' already-correct plumbing
(fastfields-cuda-impl#27) but going the other way:

• CUGLOB kernels: stride_wgt[nall] -> stride_wgt[nall+1], and compute
  wsc = stride_wgt[nall] -- the weight tensor's genuine last-dim stride
  (RLS forces this to 0 to broadcast instead).
• CUHOST launchers: copyToDevice(stride_wgt, nall) ->
  copyToDevice(stride_wgt, nall + 1), so the device-side array actually
  holds the element wsc reads.
• wsc threaded into every matvec_*_jrls/diag_*_jrls kernel call
  (including the two calls each relax_*_jrls_ makes internally).

Also bumps the kernels submodule pin to fastfields-kernels#70 (merged),
which gives the underlying _jrls kernel functions themselves the
matching wsc parameter this driver layer now passes through -- the two
changes must land together, since neither compiles against the other's
opposite side.

Status

No GPU in CI -- validated by compile+link only. Opening as draft
while a full make -C ../fastfields-cuda-lib CXX=clang++ (real nvcc, no
GPU) run finishes; will mark ready and merge once confirmed. The CPU-side
fix (fastfields-cpu-lib#65, fastfields-cpu-impl#53, fastfields-kernels#70)
is the one with a real ground-truth test run.

part of fastfields-cpu-lib#65

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com

---

Generated by Claude Code

[github-actions]

Submodule staleness (non-blocking -- fastfields-lib#15)

submodule	pinned commit	tracks	status	behind by
kernels	4333650	fastfields/fastfields-kernels@main	up to date	0

[balbasty]

Closing as superseded by #33.

A follow-up audit found that field_rls_is_jrls() (in fastfields-cpu-lib/fastfields-cuda-lib) had RLS/JRLS backwards — the opposite of the original jitfields/nitorch semantics (RLS = per-channel wc == nc, JRLS = shared/broadcast wc == 1). This PR's per-channel-stride threading into the _jrls drivers was the right shape of fix applied to the wrong (mislabelled) case, mirroring the same wrong-direction consequence of fastfields-kernels#70/fastfields-cpu-impl#53 on the CPU side.

#33 instead reverts fastfields-cuda-impl#27's wrong-direction wsc=0 force, restoring the genuine per-channel stride in the _rls drivers (no CUDA-side mirror of #70/#53 had been merged into this repo's main beyond this draft, so no kernels-pin bump was needed here). The real fix — flipping field_rls_is_jrls() — lands in fastfields-cpu-lib#73 / a matching fastfields-cuda-lib predicate flip, both closing fastfields-cpu-lib#65.

Validated by compile+link via fastfields-cuda-lib's CI (build-via-cuda-lib, real nvcc, no GPU) — green.

---

Generated by Claude Code