[CPU][ARM] Constrain MHA single-token dot_product templates (#34616)
### Details:
- Updated ARM64 `dot_product(TA*, TB*, ...)` to only participate for
float/float16 template types via SFINAE
- Updated ARM64 `dot_product(TA*, uint8_t*, ...)` to only participate
for TA in {float, float16} via SFINAE
- Kept SVE path only for `float x float` and used the NEON path as
fallback for other supported ARM64 combinations
- Wrapped bf16 query dispatch in `mha_single_token` with
`!defined(OPENVINO_ARCH_ARM64)` to avoid ARM64 bf16 instantiation paths
### Tickets:
- CVS-182514