[core] On Linux try return not used memory to system on compile model destructor

[praasz]

Details:

• Optimize Linux process using OV memory consumption by returning not used memory to the system when compiled model is destroyed.
• It should optimize memory usage by C++ and Python applications, as Linux avoid reclaim memory until process end especially for small chunks of allocations.

Tickets:

• CVS-149497

[ilya-lavrenov]

Do we need the same for other objects like ov::Model? or maybe even after some processes like compile / import model ?

[praasz]

Do we need the same for other objects like ov::Model? or maybe even after some processes like compile / import model ?

I've tried put in Model class but then is very easy to observe that execution time of loop: read, compile, infer, increased. The reason of that was that Model destructor were call many times an malloc_trim were called many times and also more system call for memory allocation occurred.
On Controlnet model from ticket the loop of 5 iterations the execution time changed from about 140s to 170s.

I would be very careful with using this function, maybe add in Core destructor is good point as it should not occur very often.
For compile / import if next actions will inference then is big chance that not used memory from these will be re-used by inference. After compile / import and some properties set the Core instance can be not required and its removal can trigger OS notification.