fix: speedup GPU kernels that raise errors

[maxymnaumchyk]

A simple benchmark comparing old CUDA kernels and new ones that use a global variable to keep track of the errors:

=== awkward_RegularArray_getitem_next_array_regularize ===
  happy path: [2, 4, 0, 2, 4] (expect [2, 4, 0, 2, 4])  old==new ✓
  error paths OK
  [.cu     ] 0.000266 s/iter
  [compute ] 0.000435 s/iter
  data: 152.6 MB,  reg_to[:4]=[500, 501, 502, 503]

=== awkward_ListOffsetArray_toRegularArray ===
  happy path: size=3 (expect 3)  old==new ✓
  error paths OK
  [.cu     ] 0.000177 s/iter
  [compute ] 0.000878 s/iter
  data: 76.3 MB,  size=5 (expect 5)

=== awkward_ListArray_validity ===
  happy path OK (no exception)  old==new ✓
  error paths OK
  [.cu     ] 0.000273 s/iter
  [compute ] 0.001328 s/iter
  data: 152.6 MB

=== awkward_ListArray_getitem_next_at ===
  at=2:  [2, 7, 12] (expect [2, 7, 12])  old==new ✓
  at=-1: [4, 9, 14] (expect [4, 9, 14])  old==new ✓
  error paths OK
  [.cu     ] 0.000429 s/iter
  [compute ] 0.000681 s/iter
  data: 228.9 MB,  tocarry[:4]=[3, 13, 23, 33] (expect [3, 13, 23, 33])

[codecov]

Codecov Report

❌ Patch coverage is 13.33333% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.46%. Comparing base (722a1b9) to head (c3f03d5).
⚠️ Report is 17 commits behind head on main.

Files with missing lines	Patch %	Lines
src/awkward/_connect/cuda/_compute.py	13.33%	26 Missing ⚠️

❌ Your patch check has failed because the patch coverage (13.33%) is below the target coverage (98.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

Files with missing lines	Coverage Δ
src/awkward/_connect/cuda/_compute.py	41.35% <13.33%> (-1.40%)	⬇️

... and 18 files with indirect coverage changes

[github-actions]

The documentation preview is ready to be viewed at http://preview.awkward-array.org.s3-website.us-east-1.amazonaws.com/PR4189