Perf: lower init peak

[loiccoyle]

I did a.second pass of cursoring memory gains, I leave this up to you if you think it is worth merging, it is not required for anything I'm working on. Feel free to close this PR if there is nothing of value :). Thanks for reviewing and getting the previous one in!

Second performance pass to get 1024 cubes through:

• Sweep the SOR update in x slabs instead of over the whole volume. Every neighbour of an active chequerboard voxel has the opposite colour, so it is never written during the same iteration and slab-wise in-place updating is exact. Bounds the increment buffer to a fixed budget instead of 4 bytes/voxel.
• Drop the two full-volume chequerboard masks and select the active colour through strided views (-2 bytes/voxel).
• Reduce ElectrodeSolver.compute_metrics from a full flux field to a chunked (batch, Nx) reduction, which was the largest transient at 8 bytes/voxel.
• Keep the conductive mask boolean and the reactive neighbour counts uint8 rather than widening both to float32.
• Compute neighbour sums directly from the unpadded volume with explicit ghost values, removing the padded copy each init made.
• Write the initial field straight into the padded buffer instead of building mask * profile separately.
• Release cached blocks around the largest init allocation; the reserved pool for a 1024^3 solve drops from 20.3 to 14.7 GiB.

Tested to be equivalent to the current implementation, and no speed decrease.

[loiccoyle]

FYI: this is what I'm getting on the benchmark

[daubners]

Very interesting, thanks! Definitely makes sense that exploiting the structure of the SOR scheme saves a lot of memory (almost -20%). I'll have a more detailed look but my current gut feeling is that we would trade a lot of readability and extensibility (from a human perspective) for performance but in a research code the first two are definitely important :D