[GPU] Fix different element types of MatMul dequantization scales issue

[yuanxion]

Details:

• MatMul dequantization Convert both dequantization scale variables (mulConst1 & mulConst2) to f32 instead of just one (mulConst2), to avoid different data type complaint issue (f16 & f32).

Tickets:

• 151988

[yuanxion]

This PR is the replacement of PR https://github.com/openvinotoolkit/openvino/pull/27075 to master.

[dmitry-gorokhov]

@v-Golubev Please take a look

[yuanxion]

May I ask you to cover the fix by functional test? I'd suggest to add a test case to src/common/low_precision_transformations/tests/mat_mul_with_constant_transformation.cpp: you can copy the existing test case, but create actual.dequantizationOnData.multiply with constant precision = f16 (you can use setConstantPrecision API for that).

@v-Golubev Thanks for your advice. Added one new testcase for it, which can be only passed when applying this PR.

(ov_env) Debug$ ./ov_lp_transformations_tests --gtest_filter=smoke_LPT/MatMulWithConstantTransformation*
[----------] 56 tests from smoke_LPT/MatMulWithConstantTransformation (1385 ms total)

[----------] Global test environment tear-down
[==========] 56 tests from 1 test suite ran. (1385 ms total)
[  PASSED  ] 56 tests.

[v-Golubev]

build_jenkins

[v-Golubev]

build_jenkins