Fix hardcoded f32 type for neg_one constant in RotaryEmbedding (#35597)
### Details:
Fixed a type mismatch issue in the RotaryEmbedding implementation where
the neg_one constant was always created with ov::element::f32,
regardless of the actual input element type. This could lead to
unexpected results or crash when the input uses a different precision
(e.g., f16). Changed neg_one to use input.get_element_type() to match
the input tensor's data type.
### Tickets:
https://jira.devtools.intel.com/browse/CVS-185685
---------
Co-authored-by: Mikhail Ryzhov <mikhail.ryzhov@intel.com>