Implement voxora-parakeet engine (NVIDIA Parakeet ASR via Rust NeMo port)

Tracks the planned `voxora-parakeet` engine crate. Per ROADMAP
Phase 7.

## Status

**Blocked on upstream.** There is currently no Rust binding for
NVIDIA NeMo (the framework that ships Parakeet). The reference
implementation is Python-only. Porting NeMo to Rust is a
multi-month project.

## What `voxora-parakeet` will be

A new workspace crate that:

- Implements `voxora_traits::AsrEngine` and
`voxora_engine::EngineAdapter` for NVIDIA Parakeet models
(parakeet-tdt, parakeet-ctc, etc.).
- Supports both `voxora-bridge::EngineFamily::Parakeet` and the
`engine-adapter` Cargo feature.
- Uses `voxora-hf` for model download (if a HF release exists for
the model; otherwise `voxora-local` is required).
- Supports `voxora-backend` device selection (cuda/metal/cpu).
- Publishes as a new crate, added to the next coordinated voxora
release.

## Upstream blockers

- [NeMo](https://github.com/NVIDIA/NeMo) has no Rust port. A
community attempt (e.g. `candle` integration) would unblock
this.
- Some Parakeet checkpoints are published on HuggingFace under
permissive licenses, but the inference code is in Python
(NeMo).

## Acceptance

- `voxora-parakeet 0.X.0` published.
- `voxora-bridge::EngineFamily::Parakeet` added (requires updating
the non-exhaustive enum in `voxora-engine`).
- Parity test against a real Parakeet checkpoint produces the
expected transcript.
- CI parity suite (#[ignore]-gated) downloads a small Parakeet
checkpoint and asserts on a known audio file.

## Out of scope

- Training or fine-tuning Parakeet models. This issue is
inference-only.