Drama speech: expose request sampling controls #1273

The native Drama engine accepts temperature, top-p, top-k and a seed, but /v1/audio/speech rejects these request fields and always submits the server's sampling constants.

Proposed contribution: accept optional request sampling controls, preserve the current behavior when fields are absent or null, validate their types/ranges before engine submission, and carry them through reference-conditioned requests, streaming and redraws. The HTTP contract should document a portable top-k range and deterministic explicit-seed retry behavior. This does not change the sampler kernels or claim parity between different backends.

Behavioral reference: Sogni-AI/drama3-mlx public generate API. The proposed Zig code will extend TensorFold's own implementation under Apache-2.0; no Fish-derived MLX source will be copied.

Target: current native main, with separate backports to the engine pins used by plotarmordev/drama-tensorfold. Tests will capture actual submitted engine requests, cover invalid numeric edge cases and preserve existing default-seed/redraw assertions. Real-model checks and any unrun backends will be reported separately from host tests.