Fix NaNs in rasterizeTopContributingGaussianIdsForward weights (#400)

I was finding that `rasterizeTopContributingGaussianIdsForward` during
training could occasionally produce NaNs in its returned weights.
Logically I don't understand how a condition like this could be reached,
but initializing the output weights to zeros (instead of
`torch::empty`), mitigates this issue.

Signed-off-by: Jonathan Swartz <[email protected]>

Author: swahtz

src/fvdb/detail/ops/gsplat/GaussianRasterizeTopContributingGaussianIds.cu

@@ -387,7 +387,7 @@ launchRasterizeTopContributingGaussianIdsForwardKernel(
         idsToRenderVec.push_back(
             torch::empty({size, settings.numDepthSamples}, means2d.options().dtype(torch::kInt32)));
         weightsToRenderVec.push_back(
-            torch::empty({size, settings.numDepthSamples},
+            torch::zeros({size, settings.numDepthSamples},
                          means2d.options().dtype(c10::CppTypeToScalarType<ScalarType>::value)));
     }