GPU perf: batched sCMOS kernels, unified retry loop, NVML scheduling

[kalidke]

Summary

• Batched sCMOS GPU kernels via KernelAbstractions with sparse coordinate extraction and in-place DoG — sCMOS GPU throughput 19 → 2,557 images/s (256×256)
• Unified GPU retry loop handling NVML polling, TOCTOU context races, and runtime OOM in a single path with jittered backoff
• BoxerConfig struct with PSF-aware and direct-control parameter interfaces, config-based calling convention
• Memory-aware batching on both CPU and GPU paths with recommend_batch_size() utility
• (ROIBatch, BoxesInfo) return tuple with timing, backend, batch, and memory metadata
• SMLMData compat bumped to 0.7 (AbstractSMLMConfig/AbstractSMLMInfo inheritance)

Breaking Changes

• getboxes now returns (ROIBatch, BoxesInfo) tuple instead of bare ROIBatch
• SMLMData 0.7 required (was 0.5)
• auto_timeout default changed from 30s to 300s

Performance (descent, RTX A6000)

Image Size	Ideal GPU	sCMOS GPU (before)	sCMOS GPU (after)
128×128	4,564	2,481	4,090
256×256	3,752	19	2,557
512×512	1,553	19	927

Test plan

• Pkg.test() passes (80 tests + benchmarks + GPU wait tests on descent)
• Documenter.jl docs build clean
• CI passes on GitHub Actions

🤖 Generated with Claude Code