fix: gpt-oss disallows quantized cache types due to its use of attention sinks

[louis-jan]

Describe Your Changes

• GPT-OSS explicitly disallows quantized cache types due to its use of attention sinks, loading this model with MLX backend now will throw quantized attention does not support zero sinks.
• QuantizedKVCache should be enabled via settings or at least detected for support before enabling by default.

Fixes Issues

• Closes #
• Closes #

Self Checklist

• Added relevant comments, esp in complex areas
• Updated docs (for bug fixes / features)
• Created issues for follow-up changes or refactoring needed

[Copilot]

Pull request overview

This PR prevents MLX inference from enabling quantized KV cache settings by default (to avoid failures with GPT-OSS models that use attention sinks), and improves streaming usage reporting by populating prompt/total token counts from MLX generation info.

Changes:

• Stop passing kvBits, kvGroupSize, and quantizedKVStart into GenerateParameters (disables default KV-cache quantization).
• Populate UsageInfo.prompt_tokens and UsageInfo.total_tokens for streaming completions using info.promptTokenCount and info.generationTokenCount.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[louis-jan]

There will be a follow-up PR to enable quantized kvcache properly.