Efficient SparseTabular construction

[WhiffleFish]

We were previously filling a dense array and then converting to sparse. This tends to be very slow for sparse problems, even yielding OOM exceptions for very large problems like Laser Tag.

Benchmarks for RockSamplePOMDP:

Pre-PR

julia> @benchmark NativeSARSOP.ModifiedSparseTabular(pomdp)
BenchmarkTools.Trial: 6351 samples with 1 evaluation.
 Range (min … max):  599.875 μs …   2.761 ms  ┊ GC (min … max): 0.00% … 76.25%
 Time  (median):     731.875 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   786.142 μs ± 284.696 μs  ┊ GC (mean ± σ):  8.76% ± 14.49%

  ▅▄▄▄▆█▆▃▁                                                     ▁
  █████████▇▄▄▃▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▃▄▆▅▇▇▇▇▇▆▅▄▃▃▄▁▆▄▄▇▆▆▇█▇█▇▇▇▇▇ █
  600 μs        Histogram: log(frequency) by time       2.11 ms <

 Memory estimate: 2.61 MiB, allocs estimate: 324.

Post-PR

julia> @benchmark NativeSARSOP.ModifiedSparseTabular(pomdp)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  173.375 μs …   2.431 ms  ┊ GC (min … max): 0.00% … 89.50%
 Time  (median):     194.583 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   215.182 μs ± 180.991 μs  ┊ GC (mean ± σ):  9.85% ± 10.33%

  █▁                                                             
  ██▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▁▂▂▂▂▂▂▂▂ ▂
  173 μs           Histogram: frequency by time         1.59 ms <

 Memory estimate: 636.45 KiB, allocs estimate: 975.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (764f4ae) 99.13% compared to head (d865c65) 99.12%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #15      +/-   ##
==========================================
- Coverage   99.13%   99.12%   -0.01%     
==========================================
  Files          13       13              
  Lines         692      689       -3     
==========================================
- Hits          686      683       -3     
  Misses          6        6              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mykelk]

This is great! Do you want to merge this @zsunberg ?

[zsunberg]

@WhiffleFish should be able to.