Investigate slow-down in reverse-mode with nested loops in Bezier curves PR

[saipraveenb25]

There appears to be some slow-down when using reverse-mode with multiple nested loops. See bezier_curvefit.py in slang-torch PR#1.

Filing this in the main repository because its most likely an issue with recompute/checkpoint policy creating unnecessary intemediates.

[saipraveenb25]

No compiler changes needed. Fixed by updating user code to take advantage of recomputation.
See for more details: shader-slang/slang-torch#6