Fix cells with duplicated indexes

[plutasnyy]

The post processing in table-transformer allows for returning multiple cells with same indexes. It happens when 1 cell is covered in more than 50% by two different spanning cells. This happens because this 'cell' is later assigned as subcell for both spanning cells instead of one with the highest probability.

Example:
Simple 2x2 table

# +-----------+----------+
# |    one    |   two    |
# |-----------+----------|
# |    three  |   four   |
# +-----------+----------+

With spanning cells over cells 'one three' (column spanning cell) and another one over 'one two' (row spanning cell). In this case cell "one" will be assigned to both spanning cells. Reproduction:

from pprint import pprint

from unstructured_inference.models.tables import structure_to_cells

table_structure = {
    "rows": [
        {"bbox": [0, 0, 10, 20]},
        {"bbox": [10, 0, 20, 20]},
    ],
    "columns": [
        {"bbox": [0, 0, 20, 10]},
        {"bbox": [0, 10, 20, 20]},
    ],
    "spanning cells": [
        {"bbox": [0, 0, 20, 10], "score": 0.9, "projected row header": False},
        {"bbox": [0, 0, 10, 20], "score": 0.8, "projected row header": False},
    ],
}
tokens = [
    {"text": "one", "bbox": [0, 0, 10, 10], "span_num": 1, "line_num": 1, "block_num": 1},
    {"text": "two", "bbox": [0, 10, 10, 20], "span_num": 1, "line_num": 1, "block_num": 1},
    {"text": "three", "bbox": [10, 0, 20, 10], "span_num": 1, "line_num": 1, "block_num": 1},
    {"text": "four", "bbox": [10, 10, 20, 20], "span_num": 1, "line_num": 1, "block_num": 1},
]

predicted_cells, _ = structure_to_cells(table_structure, tokens=tokens)
pprint(predicted_cells)

This yields:

[
....
 {'cell text': 'one three',
  'column_nums': [0],
  'row_nums': [0, 1]
},
 {''cell text': 'two',
  'column_nums': [0, 1],
  'row_nums': [0]
}]

You can see coordinates (0,0) are included in both spanning cells.

This PR fixes this by assigning only to the most probably spanning cell