Extract common pipeline from einsum2 entry points in lib.rs

[shinaoka]

Summary

strided-einsum2/src/lib.rs has 4 public einsum2 functions that duplicate the entire pipeline:

1. einsum2_into() — generic Scalar with ElementOp
2. einsum2_naive_into() — custom ScalarBase with closures
3. einsum2_with_backend_into() — custom backend dispatch
4. einsum2_into_owned() — owned arrays with conjugation

Each repeats: plan creation → dimension validation → trace reduction → permutation → fast-path check → contiguous prep → GEMM dispatch.

Proposal

Extract the common pipeline into a single internal einsum2_core() function. The 4 public functions become thin wrappers that provide type-specific parameters. Estimated ~300 lines reduction.

Risk

Medium. Requires careful trait design to avoid excessive indirection, but the benefit (single source of truth, easier maintenance) is substantial.