perf: apply RoPE in place

[vicotrbb]

What changed

• Apply rotary position encoding inside existing query and key projection buffers for single and batched inference.
• Preserve the public allocating API and add independent bit-identical parity coverage that verifies storage is not replaced.
• Correct the published crate documentation, record the allocation diagnostic, and document the required next pre-1.0 compatibility boundary.

Why

The prior internal path allocated a replacement vector for every query and key head, then copied those results into two more replacement vectors per layer. Reusing buffers already owned by the inference step removes deterministic allocator overhead without changing observable token output.

Validation

• Rust 1.96.1 formatting and strict all-target, all-feature Clippy
• Full default-feature and all-feature workspace test matrices
• Strict rustdoc and workspace doctests
• x86_64 macOS all-target, all-feature workspace check and batched-decode tests
• 16 optimized real-model Tier 1 HTTP and async OpenAI client tests with local Qwen2.5 0.5B Q4_K_M
• Exact baseline and candidate token-trace parity
• Documentation, repository layout, eval harness, release tooling, dependency audit, license, source, and unused-dependency gates
• Fixture and model crate package verification, plus inference package-content inspection

The benchmark host failed the clean-load preflight, so the recorded timing remains diagnostic and this PR makes no throughput claim.