fix(metal): coalesce adjacent --layers model map spans

[paride-cmd]

Summary

• `ds4_gpu_set_model_map_spans` now coalesces abutting/overlapping tensor spans (with page/64KiB alignment slop) before creating Metal no-copy buffers.
• A single coalesced range uses the normal overlapping `set_model_map_range` path; remaining ranges use overlapping views without the 128 GiB default span-map cap.
• This targets the `--layers` / pipeline-parallel shard case from --layers maps the shard as N disjoint Metal buffers even when the range spans the whole file (~77x decode slowdown) #845 where the "restricted" map still covered the whole shard file but as ~156 disjoint buffers (~77x decode slowdown).

Fixes #845

Test plan

• `make ds4_metal.o` on Apple Silicon
• On a Metal Flash shard with `--layers`: log should show overlapping/coalesced mapping instead of `N disjoint shared buffers across N tensor spans`
• Spot-check decode t/s recovers toward the no-`--layers` single-process baseline on the same host
• Optional: SSD-streaming / TP span maps still start (coalesce only merges abutting ranges)

AI usage disclosure: YES — assisted by Cursor agent; change reviewed against #845 measurements and existing Metal map comments.

Made with Cursor