Replace remaining uses of linearize_access_indexes for broadcasting with BroadcastIndexesRange

[swolchok]

BroadcastIndexesRange is more efficient than linearize_access_indexes-based loops. We have several remaining places that use linearize_access_indexes, and what they have in common is that they're not looping over an existing Tensor's .sizes() and .strides(). To make these work nicely, we need a way for BroadcastIndexesRange to play nicely with .sizes() and .strides() ArrayRefs that aren't attached to a Tensor.

Specific usage sites that can be improved:

executorch/kernels/portable/cpu/op_cdist_forward.cpp

Line 79 in 95f779a

x1_base_ix = linearize_access_indexes(out_base_coord, out.dim(), x1);

https://github.com/pytorch/executorch/blob/95f779ae2120d94e20bb95ae6af45da76ce3ff52/kernels/portable/cpu/op_index_put.cpp#L145.(I think?)

executorch/kernels/portable/cpu/op_split_with_sizes_copy.cpp

Line 129 in 95f779a

size_t in_linear_index = linearize_access_indexes(

executorch/kernels/portable/cpu/op_masked_select.cpp

Line 125 in 95f779a

linearize_access_indexes(broadcast_indexes, broadcast_ndim, in);

(might be straightforwardly fixable right away with no extra work)

cc @larryliu0820 @manuelcandales

[swolchok]

filed an issue because I don't necessarily intend to get to this one right away.