Support --ssd-streaming alongside --mtp / --dspark

### Feature Description
Currently, attempting to run `ds4` with both `--ssd-streaming` and `--mtp` (or `--dspark`) results in the following operational limitation/error:
> `ds4: --ssd-streaming is not compatible with --mtp yet`

Support for combining `--ssd-streaming` with `--mtp`/DSpark speculative decoding would be extremely beneficial for hardware configurations with memory constraints (such as Apple Silicon devices with 64 GB Unified Memory).

### Motivation & Use Case
On systems like a Mac Studio M1 Ultra (64 GB RAM), running large quantizations requires SSD streaming to keep the memory footprint manageable and prevent OOM / system swapping. However, streaming experts from SSD naturally introduces latency during token generation.

Combining `--ssd-streaming` with DSpark / MTP speculative execution would offer the best of both worlds:
1. Low RAM usage via dynamic expert streaming from fast NVMe SSDs.
2. Improved tokens-per-second generation speed via speculative decoding.

### System Configuration
- **Hardware:** Mac Studio M1 Ultra (64 GB RAM)
- **Tooling:** `ds4` with DeepSeek / DSpark models

Thank you for all the incredible work on `ds4`!

Start by locating ds4's argument handling and the compatibility check that rejects --ssd-streaming with --mtp or --dspark. Trace the existing streaming and speculative-decoding paths, then verify that the combined invocation works for the described DeepSeek/DSpark setup without the reported limitation or excessive memory use.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.