fix(types): set optimal default top_p threshold for nucleus sampling

[adri22235]

Description

Restores explicit default top_p=0.95 across configuration classes (GenerateContentConfig, GenerationConfig, Model, LiveConnectConfig) in google/genai/types.py.

Motivation

Empirical testing on fine-grained multimodal reasoning and deterministic DGT regulation benchmarks demonstrated that unconstrained sampling at temperature: 1.0 during truncated thinking effort causes hallucinations and sampling drift. Setting top_p=0.95 restores grounded, stable behavior.

Verification

• Verified locally on Debian 12 (Linux x86_64).
• Tested and built clean wheel package (.whl) with 0 errors.
• CLA signed by author (@adri22235).