[NNPA] Support matching tensor information in the JSON config file for NNPA

[tungld]

This PR supports matching tensor information such as: rank, data type, dimension size in the JSON config file for NNPA.

For example: match MatMul operations of A_3D x B_2D and the last dimensions of A and B are divisible by 64, then use NNPA for those MatMul operations.

{
  "nnpa_ops_config": [
    {
      "pattern": {
        "match": {
          "node_type": "onnx.MatMul",
          "inputs": {
            "0": {
              "rank": "3",
              "type": "f32",
              "dims": {
                "-1": "%64==0"
              }
            },
            "1": {
              "rank": "2",
              "type": "f32",
              "dims": {
                "-1": "%64==0"
              }
            }
          }
        },
        "rewrite": {
          "device": "nnpa"
        }
      }
    }
  ]
}

Supported patterns for checking integers (rank and dimension size):

Comparison Operators:

• "3" - Exact match (implicit equality): value must equal 3
• ">3" - Greater than: value must be > 3
• ">=3" - Greater than or equal: value must be >= 3
• "<3" - Less than: value must be < 3
• "<=3" - Less than or equal: value must be <= 3
• "==3" - Explicit equality: value must equal 3
• "!=3" - Not equal: value must not equal 3

Modulo Operations (for divisibility/alignment checks):

• "%32==0" - Modulo constraint: (value % 32) must equal 0
• "%64==0" - Divisibility by 64: (value % 64) must equal 0
• "%N==R" - General form: (value % N) must equal R

Input/output/dim index

• Support negative index, for example, -1 is the last input/output/dim, -2 second to the last, etc.

See the documents JsonConfigFile-NNPA.md for more examples.