test: add speculative decoding E2E test to CI pipeline

[solderzzc]

Close #24

Summary

This PR finalizes the integration of the SSD expert streaming and speculative decoding pipeline into the SwiftLM production codebase, and adds continuous integration testing for dual-model verification.

Key Changes

1. Dependency Sync: Updated Package.swift to point the mlx-swift-lm dependency to SharpAI/mlx-swift-lm branch main, tracking the integrated PR Feature/gemma4 benchmark #9 which contains the SSD streaming 10x rewrite.
2. README Documentation:
   • Added a new SSD Expert Streaming: 10x MoE Speedup section detailing the methodology (cross-projection batching, concurrent pread, asyncEval pipeline, persistent Metal buffers, runtime top-k).
   • Included full benchmark results showing 0.58 → 6.52 tok/s improvement for 122B+ MoE models on M1 Ultra 64GB.
   • Updated CLI options table to include --draft-model and --num-draft-tokens.
   • Credited Eric Lake in the Acknowledgments for the speculative decoding infrastructure and SSD rewrite.
3. Speculative Decoding CI:
   • Added tests/test-speculative.sh: A new dual-model E2E integration test verifying speculative decoding path activation, sequential stability, memory limits, and streaming behavior.
   • Updated .github/workflows/ci.yml: Created a new speculative-decoding job leveraging the macos-15-xlarge (14 GB) runner, using Qwen3.5-0.8B-MLX-4bit as a lightweight draft model to accelerate a Qwen3.5-9B-4bit main model.

Notes for Reviewers

• The new CI job is constrained to macos-15-xlarge because loading both the 0.8B and 9B models requires ~6GB resident RAM, which is too tight for the standard 7GB macos-15 runner without risking OOM terminations.
• The test-speculative.sh intentionally verifies the presence of the Using speculative decoding logs to ensure the draft model is actively engaged in the generation pipeline.