perf: reduce memory usage

[loiccoyle]

Summary

Reduces GPU memory use across TauFactor solvers while preserving benchmarked tortuosity values and convergence iteration counts.

• Stores segmentation labels as uint8; uses boolean checkerboards/masks.
• Releases initialization temporaries earlier and avoids unnecessary full-volume field-profile repeat() allocations.
• Reuses a solve-local increment buffer and accumulates neighbour sums in-place.
• Replaces full-volume rolling neighbour sums with direct slice-based sums, including periodic Y/Z wrapping.
• Extends benchmark output with separate initialization and solve RAM metrics: VRAM_I_max, VRAM_I_cur, VRAM_S_max, and VRAM_S_cur.

NVIDIA RTX A6000 benchmark

Compared fcc structure runs up to 1024³ for current main versus this code, using VRAM(max) and solver wall time :

Solver	Peak VRAM change	Solve-time change
Solver	−32.2%	+4.7%
PeriodicSolver	−32.1%	+4.6%
ElectrodeSolver	−21.5%	+4.7%
PeriodicElectrodeSolver	−21.4%	+4.6%

Before these changes, initialization was particularly limiting for periodic solvers: periodic electrode initialization peaked at 424 MB versus 338 MB for the non-periodic version. Direct periodic neighbour sums remove that rolling-allocation bottleneck; both now peak at 287 MB during initialization. Overall peak memory is now determined by the solve phase rather than construction.

Solver	Init peak reduction	Init current reduction
Solver	−19.9%	−28.3%
PeriodicSolver	−24.0%	−28.5%
ElectrodeSolver	−14.9%	−28.3%
PeriodicElectrodeSolver	−32.2%	−28.5%

Validation

• Benchmark tau values and iteration counts are unchanged in the compared runs.
• Focused solver and benchmark tests pass; lint and diff checks pass.

[loiccoyle]

🎉