chore(release): refresh Linux aarch64 CUDA artifact for specialized serving

## Problem

The official v0.6.0 Linux aarch64 CUDA 13 archive is older than the specialized-serving behavior now required by the public recipes catalog:

- Text-generation recipes for Qwen, Gemma, and Llama reach the chat-completions boundary on an NVIDIA GB10.
- The archive does not expose the current dedicated embedding-serving interface needed by the Qwen3 Embedding recipe.
- A Whisper checkpoint loads and serves the audio transcription route, but `/health` continues to inspect the unavailable chat worker and remains at HTTP 503.

A current-main CUDA build compiled for SM 12.1 passes all five seed recipes through their task-specific HTTP endpoints on the same GB10. That proves the source behavior, but it must not be published as verification of the older v0.6.0 archive.

## Proposed work

- Publish a new versioned Linux aarch64 CUDA 13 release archive after the specialized embedding and audio-health changes reach the public repository.
- Exercise the exact downloaded archive rather than an in-tree build.
- Feed the resulting rendering hashes and logs into the downstream recipes verified-badge workflow.

Do not replace the existing v0.6.0 asset in place; a new version keeps download provenance and cached artifacts unambiguous.

## Acceptance criteria

- [ ] The new archive contains aarch64 ELF binaries dynamically linked to the expected CUDA 13 libraries with no missing shared objects on GB10.
- [ ] `mlxcel-server` from that archive starts a Qwen3 embedding checkpoint in dedicated embedding mode and `/v1/embeddings` returns a non-empty vector.
- [ ] `mlxcel-server` from that archive starts Whisper, `/health` reaches HTTP 200 after load, and `/v1/audio/transcriptions` returns non-empty text.
- [ ] Qwen3 4B, Gemma 4 E4B, and Llama 3.1 8B chat recipes return a non-empty `content` or `reasoning_content` response.
- [ ] The exact release version and binary checksum are preserved with the five-seed GB10 verification evidence.

Use the exact downloaded versioned archive rather than an in-tree build. Start with the mlxcel-server entry point and exercise the five seed recipes through their listed HTTP endpoints on a GB10, checking binaries, shared libraries, responses, health, logs, rendering hashes, and the checksum. Done means a new archive passes every acceptance criterion with its version and verification evidence preserved.

Written by the indexing model from the issue text.