Fix ReadValue variable memory by using fake alignment layout

[yuanxion]

Details:

Use actual size of ReadValue variable memory to pass the ocl::check_boundaries checking during gpu_usm::copy_from().

Tickets:

• CVS-165646

[yuanxion]

For the florence2 notebook, the ReadValue node's variable memory layout 3x590x768 is smaller than the fake aligned memory layout 1776x1x768 (equals to 3x592x768) of its dependency FC node, and cannot pass the ocl::check_boundaries checking.

[yuanxion]

It can pass the ocl::check_boundaries checking and run florence2 models successfully after applying the commit f3e1415.
Here is the object detection result:

However, this result is a little different with the result that does not use this PR and just comment out the ocl::check_boundaries checking, which has a door detected (need to check the diffenence later):

[yuanxion]

Reverted previously fake aligned memory size for ReadValue, and just use the size of ReadValue variable's memory when copy_from its dependency memory.

[yeonbok]

Could you add a func test?

[yuanxion]

Could you add a func test?

@yeonbok, @p-durandin Done, testcase is added.

Note: Google Test filter = variable_test_common.variable_copy_from_fc
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from variable_test_common
[ RUN      ] variable_test_common.variable_copy_from_fc
[       OK ] variable_test_common.variable_copy_from_fc (343 ms)
[----------] 1 test from variable_test_common (343 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test suite ran. (344 ms total)
[  PASSED  ] 1 test.