FIX: CUDA misaligned address (716) error by avoiding merged return in BSphere construction from triangle vertices

[fjbarter]

Rather than constructing centre and radius in the branches, and merging before returning, just return inside the branches.

Changed src/bounding_volumes/bsphere.jl

Stella's examples/rotating_drum/rotating_drum.jl now works (05/04/2026) with PrecisionSettings(CUDA) with both single and double precision

Benchmarked and proven equivalent performance:

Merged return:
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  105.800 μs …  3.796 ms  ┊ GC (min … max): 0.00% … 0.00%
  ▃▅▃▂▃▃▄▅▅▇███████▆▇███▆▅▅▄▄▃▂▂▂▂▂▂▂▁▁▁▁▂▄▄▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▃
  106 μs          Histogram: frequency by time          168 μs <

 Memory estimate: 704 bytes, allocs estimate: 44.
Early return:
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  105.900 μs …  3.403 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     122.600 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   126.722 μs ± 56.277 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

            ▂█▂▁      ▁▁
  ▂▃▂▂▂▂▃▃▄▄████▇▇▇▆▅▅██▇▅▄▅▄▃▃▂▂▂▁▁▂▁▁▁▁▁▁▂▅▄▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▂
  106 μs          Histogram: frequency by time          166 μs <

 Memory estimate: 208 bytes, allocs estimate: 13.
Median merged: 0.122 ms
Median early : 0.123 ms

[anicusan]

This is great, thanks Freddie!