`gather` should allow 0D output

[Honry]

The spec allows a scalar (0-D) indices operand. Since output rank = rank(input) + rank(indices) − 1, a scalar index over a 1-D input yields a 0-D output:

input   : [4]   (rank 1)
indices : []    (rank 0, scalar)
axis    : 0
=> output: []   (rank 0)

But gather's output rankRange.min is currently 1, which rejects this valid case. The minimum output rank should be 0.

[Honry]

cc/ @fdwr

[fdwr]

Indeed, output ranges 0 through input.rank() + indices.rank() - 1.

operand	allowed data types	required data types	allowed ranks	required ranks
input	any	"float32", "float16", "int32", "int8", "uint8"	1 to any N	1 to 5
indices	"int32", "uint32", "int64"	"int32"	any N	0 to 5
output	same as input	"float32", "float16", "int32", "int8", "uint8"	✏️0 to any N	✏️0 to 5

Dimension concatenation logic:

input.shape   = [leading dimensions..., input axis dimension, trailing dimensions...]
indices.shape = [                       index dimensions...                         ]
output.shape  = [leading dimensions..., index dimensions...,  trailing dimensions...]
output.shape  = input.shape[0..axis] ~ indices.shape ~ input.shape[axis+1..input.rank]
axis = from 0 to (input.rank - 1)
indices[∀ coordinates] < input.shape[axis]

I have an old WindowsML test case for this:

    // 0D, axis 0, rank 0 indices tensor
    {
      "op_type": "Gather",
      "axis": 0,
      "data": [1,2,3],
      "indices": 0, // 0D scalar
      "output": 1,  // 0D scalar
      "T": "float32"
    },