Expose RoPE theta via config API; cache fp32 cos/sin tables

[KakaruHayate]

Summary

Two changes in one PR:

1. Expose RoPE theta via config API — forward rope_theta (plus existing rope_interleaved) through the encoder, so RoPE theta is config-tunable instead of hardcoded to 10000.
2. Refactor RoPE to cache fp32 cos/sin tables (modules/commons/rotary_embedding_torch.py), per review comments:
   • cache fp32 cos/sin at init instead of raw angles, so trig runs once instead of every forward;
   • tables stay fp32, only cast to the activation dtype at use — removes the error of recomputing trig on fp16/bf16 angles during 16-mixed training;
   • the interleaved path no longer needs einops, simplifying exported ONNX graphs (no Cos/Sin/Einsum ops).

Compatibility

fp32 path is bit-identical to the previous implementation; only the fp16/bf16 column improves in precision. Buffers remain persistent=False, so no checkpoint or config changes — existing models resume seamlessly. No tests included, per repo convention (verified via normal community testing).