Add YaRN RoPE scaling and correct dynamic scaling

[dkuku]

YaRN checkpoints currently cannot be loaded, and dynamic RoPE raises the entire base to the scaling exponent instead of applying the exponent only to its multiplier. This adds YaRN and corrects dynamic scaling to use the highest position ID rather than the padded/cache allocation length.

• Support YaRN frequency blending, attention scaling, and truncation, with documented defaults for direct configuration.
• Validate RoPE configuration and type aliases while preserving legacy Phi-3 LongRoPE aliases.
• Add regression tests for the formulas, malformed options, and the documented YaRN configuration.

Validation: 38 targeted tests passed on this branch independently, covering layer/configuration tests and existing Llama, Mistral, Qwen3, Gemma3Text, and Phi3 tests. Formatting, compilation with --warnings-as-errors, and unused-dependency checks pass on the complete stack. A dependent embedding PR adds end-to-end Python Transformers parity fixtures for the RoPE variants.

The full test suite was not run. No associated issue is being closed.

AI disclosure: this contribution was developed with substantial LLM/coding-agent assistance, including implementation, tests, and review. Validation above was executed by the coding agent.