Split long non-causal SDPA queries to avoid Ultra corruption

[quantz8a]

Summary

• Work around MPSGraph scaledDotProductAttention silent corruption on M2/M3 Ultra for long non-causal sequences (Corrupted video (and audio) outputs #16 / Apple FB24605554).
• Chunk query rows (default 2048) when sequence > 12288; leave smaller jobs on the fast single-op path.
• Tunables: --sdpa-query-block N, H3_SDPA_MAX_QUERY_ROWS (0 disables), H3_SDPA_SPLIT_MIN_ROWS.

Fixes #16
Related: #15 (same size class; assert vs silent garbage)

Notes

• Non-causal only (DiT path). Causal GQA unchanged.
• Split can increase MPSGraph intermediate residency; combine with --ssd-streaming for long HQ jobs (as discussed in Corrupted video (and audio) outputs #16).

Test plan

• Compiles h3_gpu.m / main.c with stock flags
• make -j && ./h3_tests
• On M2/M3 Ultra: 896×512 or 1344×768, ≥100 frames, steps≥4 previously produced static → now coherent scene/audio
• Short job (512×512, 22–107f) unchanged wall-clock path (no split below gate)
• Local M3 Ultra validation in progress (896×512, 107f, steps=4, --ssd-streaming)

[quantz8a]

Local M3 Ultra (96GB) validation for this PR:

• 896×512, frames=90, steps=4, layers=45, --ssd-streaming, H3_DIT_COMMAND_BLOCKS=5
• Completed exit=0; wrote a playable MP4 (duration=3.75s)
• Mid-frame luma contrast check: mean≈17.5, std≈17.5 (not collapsed static)

This geometry is in the Ultra corruption zone discussed in #16; output is coherent rather than noise.