[TRANSFORMATIONS] SDPAToPA remove only SDPA-related Assign nodes (#33696)
### Details:
SDPAToPA aggressively removes all Assign nodes without checking what
part of the model they belong to. This has been a right thing to do
before the advent of hybrid models like LiquidAI/LFM2-1.2B that use
ReadValue/Assign not only for full attention, but also for linear
attention (not SDPA). The removal of all Assign nodes caused some
ReadValue non-related to SDPA to lose their pair which lead to a
corrupted model. Fix it by detecting and deleting only the proper Assign
nodes.
Minor refactoring to remove unnecessary parts and make the code of the
transformation simpler.
### Tickets:
- [CVS-179213](https://jira.devtools.intel.com/browse/CVS-179213)
Signed-off-by: Andrii Staikov
---------
Co-authored-by: Mikhail Ryzhov <mikhail.ryzhov@intel.com>