fix: treat DLTensor.strides == NULL as compact row-major

[balbasty]

What

DLPack permits DLTensor.strides == NULL for a compact row-major tensor, but every cpu-lib module fed x.strides straight into CANUSE32BITS / copy_if_needed and the impl loops — all of which dereference it, so a producer that omits strides would segfault.

• autocast.h: new ContiguousStrides RAII wrapper holding a normalised copy of the (POD) descriptor; when strides are NULL it synthesises explicit contiguous strides (in elements) that outlive the call. An enable flag skips optional placeholder tensors (posdef weight, mesh nearest_vertex) whose shape may be invalid and whose strides are never read.
• Every public entry (distance/posdef/resize/restrict/splinc/pushpull/reg_field/reg_flow) normalises its tensors through the wrapper before dispatch; the bodies are otherwise unchanged, so the strides-supplied path is byte-for-byte identical.
• tests/test_resize: a null-strides regression — strides=NULL must match the same call with explicit contiguous strides.

Why

nanobind always supplies strides, so the current bindings never hit this — but the library's "accepts any __dlpack__ object" contract means an external DLPack producer can.

Verified

All 8 module test suites pass with the change compiled in (each test binary links all 8 module .cpp): distance 2350, mesh 4620, spline 510, posdef 3580, pushpull 175, reg_field 272, reg_flow 282, resize 211 (incl. the new null-strides case).

Independent of the kernels half (fastfields-kernels#…, a defense-in-depth guard in canUse32BitIndexMath) — this PR supplies non-null strides so it never relies on that guard.

Part of #6.

🤖 Generated with Claude Code

---

Generated by Claude Code

[balbasty]

ok