[CORE][CPU] Add erfinv operation (#34823)
### Details:
Introduces `ov::op::v17::ErfInv` — the element-wise inverse error
function — as a new operation in OpenVINO opset17, aligned with
[`torch.erfinv`](https://docs.pytorch.org/docs/stable/generated/torch.erfinv.html).
Includes a full CPU plugin implementation with JIT emitters for x64,
aarch64, and riscv64.
### Motivation
`erfinv` appears in statistical models and normalizing flows (e.g.,
Box-Muller transform, quantile functions). PyTorch models using
`torch.erfinv` cannot currently be converted to OpenVINO without a
workaround. This PR adds native support so such models convert and run
efficiently without falling back to a slow scalar path.
### Changes
#### Core (`src/core/`)
- New op class `ov::op::v17::ErfInv` (unary, float-only:
f16/bf16/f32/f64)
- Registered in `opset17_tbl.hpp` and `ops.hpp` / `ops_decl.hpp`
- Reference kernel (`reference/erfinv.hpp`) using the [Giles
(2010)](https://people.maths.ox.ac.uk/gilesm/files/gems_erfinv.pdf)
piecewise polynomial approximation — two 8-term Horner branches split on
`w = -log(1 - x²) ≥ 5`
### Notes
- The Giles approximation has ~6 ULP max error for f32, consistent with
PyTorch's own implementation.
### Tickets:
- 183678
### AI Assistance:
- *AI assistance used: yes
- The entire PR was generated using AI agent. Code was manually
reviewed.
---------
Co-authored-by: Katarzyna Mitrus <katarzyna.mitrus@intel.com>
