[NPUW] INT-8 dynamically quantized kv-cache (#34516)
### Details:
- Mixed kv-cache compression (key:u8-assym, values: u4-sym) might be
enabled by config option ``` "KV_CACHE_PRECISION": "u8" ```
- Currently number of compilation issue for i4 compression exists - so
default scheme is (key:u8-assym, value: i8-sym)
- Performance concerns are exist, but e2e pipeline functionally worked
fine in u8 kv-cache mode, accuracy is on pare on limited scope tests
- For value-cache experimental mode is possible : i4 - symmetrical
quantization, this verified by RMSE metric on real kv-caches, also some
work is here: [Quantize What Counts: More For Keys, Less For
Values](https://arxiv.org/abs/2502.15075v3)
### Tickets:
- C-180875
### AI Assistance:
- *AI assistance used: yes*
- *AI used to synthesize tests, and for reverse crafted decomposition v3
based on fuse-pattern*
