Support higher axis numbers for dr.any and dr.all reductions

[zhestyatsky]

Hello,

I have been experimenting with the latest drjit version and noticed that dr.all() and dr.any() operations seem to not be supported for Tensors when they are evaluated with axis argument greater than 0.

Reproducer below raises `RuntimeError: drjit.all(<drjit.cuda.TensorXb>): tensor type is not compatible with the requested reduction.

import drjit as dr
from drjit.cuda import TensorXb

v = TensorXb([[True, False, True],
              [True, True, True],
              [True, False, True],
              [True, True, True]
             ])
dr.all(v, axis=1)

Similar exception is raised when calling dr.any().

This PR aims to enable dr.all() and dr.any() operations for Tensors with axis parameter greater than 0.

[wjakob]

Very minor. rendceOp -> reduce_op (underscore case for vriables)

[wjakob]

Hi @zhestyatsky -- is this PR ready to be merged or still under development? It states "Draft" status.
One thing I am surprised about is that this works as-is.

The all/and reductions are AFAIK only implemented for 32-bit integers. We previously got away with using them for reduction down to 1 element by applying a simple follow-up 4->1 element reduction and doing something some extra memsets in case the input array wasn't exactly a multiple of 4 elements.

For reductions along another axes, I am wondering if there won't be mix-ups when the element count isn't a multiple of 4?