deps: bump kernels pin to the diag_bending/diag_all corner cross-term fix

[balbasty]

What

Moves the kernels gitlink from 3e38c85 to 1df9fd3 (current fastfields-kernelsmain), putting the CUDA path on the same kernels commit the CPU path now uses. Gitlink-only — exactly one modified path.

Why: a real, live pin skew

Tracing the actual gitlinks on main before this PR:

CPU path : cpu-lib -> cpu-impl 41d66b3 -> kernels be7be08
CUDA path: cuda-lib -> cuda-impl a099a04 -> kernels b09b284

The two backends resolved to different kernels commits, with CUDA three behind (b09b284..be7be08 = kernels #74, #76, #78). This repo's own main was in a third state again (3e38c85), because cuda-lib's impl pin is itself well behind cuda-implmain.

This PR fixes the lower half of that chain. The cuda-lib → cuda-impl half follows in a separate PR.

What it picks up here (3e38c85 → 1df9fd3)

kernels PR	commit	effect
#78	be7be08	drops the pointless virtual destructors from distance/meshNode/BoundingSphere
#79	93a954b	restrict.h includes batch.h for index2sub
#81	1df9fd3	correctness. The doubled fx1*fy0 / dropped fx0*fy1 corner term in the boundary-corrected diagonal of diag_bending (field 2d/3d, flow 2d/3d) and diag_all (flow 2d/3d). These are shared kernel headers, so the CUDA backend carried the identical defect.

3e38c85 is an ancestor of 1df9fd3, so this is a fast-forward of the pin.

Verification

Per this project's standard: the shared kernel math is validated by the CPU suite (the same headers compile for both backends), and the CUDA-only glue by nvcc compile+link, since there is no GPU in CI.

• Shared math — full CPU suite over kernels 1df9fd3: 13/13 suites, 0 failures (test_reg_field 18284 checks, test_reg_flow 11415). Additionally, the new boundary-symmetry regression tests fail loudly against the pre-fix kernels (328 + 1696 failures) and pass against 1df9fd3, so the corrected corner term is genuinely exercised.
• CUDA compile+link — this PR's build-via-cuda-lib job, which runs the repo's real invocation (make -C . -j2 CXX=clang++ CXXFLAGS="-std=c++14 -O1"), preserving the -O1 host-code setting and cuda-lib's split MODULES list.

Pin hygiene

1df9fd3 is the squash-merge commit that actually exists on fastfields-kernelsmain, not a pre-merge branch head. Verified it resolves before opening.

🤖 Generated with Claude Code

https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z

---

Generated by Claude Code

[github-actions]

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

submodule	pinned commit	tracks	status	behind by
kernels	1df9fd3	fastfields/fastfields-kernels@main	up to date	0