perf: add C-contiguous fast path to toBuffer() eliminating per-call allocation - #35

This PR adds a C-contiguous fast path to toBuffer() (and intoBuffer()) that copies data directly from the underlying Rust buffer via memcpy, skipping the intermediate full-array allocation and copy that extract_array_* incurred on every call.

Motivation and Context

fromBuffer() has always been fast: one bulk memcpy into internal storage. toBuffer() was significantly slower and degraded as array sizes grew, making it a bottleneck in ML inference pipelines (e.g. ONNX runtime output buffers, growing KV caches).

Root cause: every toBuffer() call ran extract_array_* which allocated a new Vec and copied the entire array — even when the data was already C-contiguous and could be read in-place. The subsequent copy_array_to_buffer also walked elements one at a time via stride iteration instead of using a bulk slice copy.

What's Changed

• Added C-contiguous fast path: contiguous arrays copy directly from the internal buffer (0 allocations, 1 memcpy — same cost profile as fromBuffer())
• Non-C-contiguous views (sliced, transposed, strided) fall through to the existing extract_array_* path
• copy_array_to_buffer now uses as_slice() + copy_from_slice instead of element-by-element stride iteration

Breaking Changes

None