Warn when speculative decoding may hurt throughput for MoE models

[Shylin26]

Fixes #1132

Problem

When using --draft-model with MoE target models (e.g. Qwen3.5-397B-A17B),
speculative decoding can hurt throughput by 25-45% because the target model's
active parameter count is close to the draft model size, making draft+verify
more expensive than direct generation.

Solution

Added _warn_speculative_moe() that runs after both models are loaded. It:

• Detects MoE architecture via num_experts and num_experts_per_tok in model config
• Calculates active parameters as (num_experts_per_tok / num_experts) * total_params
• Warns if draft model size is within 4x of active parameters

Non-MoE models are completely unaffected.