Align ConvertPrecision with LSB-first nibble packing for u4/i4 types

[evkotov]

This PR fixes the nibble ordering inconsistency in 4-bit types (u4/i4) where ConvertPrecision transformation was not aligned with the current LSB-first packing standard.

Problem

ConvertPrecision transformation was using deprecated MSB-first nibble order (not updated after PR #20371), while Convert operation with ConstantFolding used the correct LSB-first order via element::iterator. This created inconsistencies between
different code paths for the same input.

Path inconsistency before fix:

  Input: u4 constant bytes [0xAB]
               |
               v
        ┌──────┴──────┐
        |             |
        v             v
    ConvertPrecision  Convert → ConstantFolding
    (MSB-first)       (LSB-first via is_lsb_packed)
        |             |
        v             v
     [10, 11]      [11, 10]  <-- Different results!

This led to incorrect values in INT4 quantized models (like Phi-3) when different optimization paths were applied.

Solution

Refactored convert_low_precisions_int() to use Convert::evaluate() instead of custom conversion logic. This ensures consistent LSB-first nibble packing across all conversion paths, aligning with the standard established in PR #20371.

Path consistency after fix:

 Path consistency after fix:
  Input: u4 constant bytes [0xAB]
               |
               v
        ┌──────┴──────┐
        |             |
        v             v
    ConvertPrecision  Convert → ConstantFolding
    (via Convert::evaluate)   (LSB-first)
        |             |
        v             v
     [11, 10]      [11, 10]  <-- Same results!

Tickets:

• CVS-161778