CUDA stream: widen `int`→`intptr_t` across layers and forward it in the distance launchers

[balbasty]

Confirmed by the fable review; GPU-only, so tracked rather than fixed in the correctness pass (it's a pervasive cross-layer signature change that needs a full CPU+CUDA recompile to verify, and can't be runtime-checked without a GPU).

(a) Width. Every public stream parameter is int across fastfields-lib, fastfields-cpu-lib, fastfields-cuda-lib, fastfields-cuda-impl and fastfields-bind-py/src/ext.cpp, but a cudaStream_t is a pointer (64-bit). A real stream handle passed through int truncates. Widen to intptr_t/int64_t consistently through all layers (CPU ignores it, so no CPU impact) before any GPU wheel ships.

(b) Forwarding. In fastfields-cuda-impl/distance_euclidean.h, distance_l1.h, distance_mesh.h the kernels launch with <<<..., 0>>> and use synchronous cudaMemcpy; cuda-lib/distance.cpp marks its stream param unused and the impl dt() signatures have no stream parameter at all. Thread the stream through dt() and launch on (cudaStream_t)stream with cudaMemcpyAsync, matching the posdef/resize/restrict/splinc launchers which already forward it. (Do (a) and (b) together.)

Workstream: claude-jitfields-to-fastfields