[teeny] distance/l1.h: rewrite the L1 line kernel on the now-blessed two-pass scan_ idiom

[balbasty]

Agent: claude-fastfields-to-teeny

Part of the teeny-refactor umbrella (fastfields-lib#21). L1 distance's outer batch/axis loop is already teenified; the per-line min-plus recurrence itself is still a bespoke hand-written sweep — the last one in the module.

teeny#464/#466 (merged) just blessed exactly this shape as the documented idiom for two-pass line recurrences (forward + backward), with a worked min-plus example in teeny's own docs:

constexpr auto ax = axis<-1>{};
scan_(t, inf, minplus{w}, ax); // forward
scan_(t.flip(ax), inf, minplus{w}, ax); // backward

where minplus is a tiny device-safe functor: e = c = min(e, c + w) (matching the project's lambda-free engine convention).

Task

Replace distance/l1.h's hand-written line kernel with this pair. No fused scan2_ exists or is planned (teeny deliberately rejected it — "Keep it teeny", no profiling case for the fused form) — two plain scan_ calls is the correct, final shape, not an interim one.

Verify bit-exact against the existing L1 oracle test (tests/test_distance_l1.cpp — teeny's own PR #466 already added two cases doing this same substitution against teeny's own jitfields reference, as a cross-check pattern to follow here too).

Euclidean's line kernel (Felzenszwalb lower-envelope-of-parabolas) is not in scope here — it needs random access into whole-line scratch buffers and a data-dependent stack pop, correctly identified as not scan-shaped in the earlier deep review. Leave it as the bespoke kernel it already is.

Priority

Low — small, well-scoped, mechanical once the teeny pin is bumped past dcd591c to include #465/#466. Good candidate to bundle with whatever PR next bumps the external/teeny pin.