Add TurboQuant KV cache backend

[timonharz]

Summary

This PR ports the TurboQuant KV-cache backend from mlx-vlm PR #858 into mlx-swift-lm, with the implementation centered in MLXLMCommon so it works for both MLXLLM and the text-decoder side of MLXVLM.

The behavior matches the upstream intent:

• fractional kvBits automatically use TurboQuant
• integer kvBits continue to use uniform quantization by default
• kvQuantizationScheme = .turboQuant can explicitly force TurboQuant for integer bit widths

What changed

• Added a new shared TurboQuantKVCache backend in MLXLMCommon
• Ported the active TurboQuant codec/runtime stack, including split-codec support for .5 bit widths
• Extended GenerateParameters:
  • kvBits is now Float?
  • added KVQuantizationScheme
  • added kvQuantizationScheme
• Updated dynamic KV-cache quantization to recurse into nested CacheList contents
• Preserved existing skips for unsupported cache types such as RotatingKVCache, MambaCache, and plain ArraysCache
• Integrated TurboQuant into the shared attention/cache-update path
• Patched model-specific attention implementations that bypass the shared helper, including GPTOSS and MiMoV2Flash
• Kept the existing “no attention sinks on quantized caches” behavior for TurboQuant as well
• Extended prompt-cache save/load to serialize and restore TurboQuantKVCache
• Updated docs/examples to cover fractional kvBits and explicit quantization scheme selection

Tests

Added coverage for:

• TurboQuant codec behavior
• cache selection for fractional bits and explicit .turboQuant
• nested cache quantization behavior
• prompt-cache round-tripping for TurboQuantKVCache

[N1k1tung]

Runtime test execution is still blocked in this environment because MLX fails to load default.metallib, so I could not complete the full TurboQuant test run here.

kinda explains it was in codex sandbox =)