Fix product reduction gradient when an element is zero

[lnuic]

The gradient of a product should be the product of the other entries, but we
computed it as result / value, which breaks when an entry is zero.
dr.prod caught the divide by zero and returned 0, which is the wrong answer: for
[2, 0, 4] the gradient should be [0, 8, 0] and previous result was [0, 0, 0].

The axis/block version didn't check at all and returned NaN.

The cost is two extra reductions in the backward pass, so a product backward is about 3x slower:

backend	op	N	before	after
LLVM	prod	16M	3.8 ms	13.0 ms
LLVM	block_reduce	16M	6.0 ms	13.0 ms
CUDA	prod	16M	0.044 ms	0.135 ms
CUDA	block_reduce	16M	0.090 ms	0.314 ms

That could be brought back to a single pass by combining the two reductions, but
it needs a new reduction primitive in drjit-core, which felt like too much for
this one fix. I am happy to add that if needed.

If the slowdown isn't worth it I am happy to leave the gradient approximate and
just guard the block version so it returns 0 instead of NaN.

[wjakob]

Hi Lovro -- are you sure about needing new backend operations? Dr.Jit supports block_prefix_reduce which seems to be just what is needed. What I don't like about this version is that it seems to make the primal slower even when no derivatives are ultimately needed.