optimization ConstantFolding memory consumption with periodic cleanup

[evkotov]

Details:

OpenVINO ConstantFolding transformation is consuming a lot of memory during model compilation, particularly for large models like GPT2. Memory profiling revealed that ConstantFolding alone consumed ~500MB during transformation pipeline,
contributing to peak memory usage of over 3GB.

Memory profiling identified that ConstantFolding creates large intermediate constant nodes. These nodes are freed by the transformation logic but memory is not immediately returned to the OS due to glibc malloc
behavior - freed memory is kept in internal pools for potential reuse.

• Original ConstantFolding: ~500MB consumption (418.5MB + 81.93MB in major calls)
• Peak memory without ConstantFolding: 1897.7MB
• Peak memory with ConstantFolding: 3030.9MB
• Memory pressure caused system-level performance degradation

Implemented periodic memory cleanup in ConstantFolding transformation using malloc_trim() to force return of unused memory to OS.

Tickets: