[GPU] Resolve perf regression of sdpa_opt usage with systolic arch (#33679)
### Description of the issue
* Performance drop is introduced from OV25.2 to OV25.4 on LNL.
* SDPA caused perf drop
#### The code and line that caused this issue
- SDPAFusion is processed and sdpa layers selects sdpa_opt kernel by
supports_micro_sdpa :
sdpa_opt.cpp:229
#### Reproduction step and snapshot
- Reproduced by benchmark
`.\benchmark_app.exe -m unet_int8.xml -d GPU -t 5 -hint latency`
#### How to fix
- Not to use sdpa using SDPADecomposition if support_immad is true and
sdpa_opt kernel would be selected.
- But this exception applied only if can check memory size is
decomposable compared to max memory allocation
#### Checklist
- [x] Is it a proper fix?
- [ ] Did you include test case for this fix, if necessary?
- [x] Did you review existing test that can be extended to cover this
scenario?
- relevant test cases comparing with a reference code exists :
`INSTANTIATE_TEST_SUITE_P(smoke_ScaledAttnStatic3D_GPU,
ScaledAttnLayerGPUTest,
static_shape_params_3D,
ScaledAttnLayerGPUTest::getTestCaseName);`
### Tickets:
- CVS-178906
---------
Signed-off-by: Min, Byungil <byungil.min@intel.com>