[Feature Request] Support for Q4 / Q5 / Q6 quantizations (e.g. Unsloth's UD-Q4_K_XL to UD-Q6_K_XL) #222

Feature Description

Currently, Strata primarily supports Q2_0, IQ2_XS, IQ3_XXS, and IQ3_S quantizations to fit within standard consumer setups (~64GB RAM + 12-24GB VRAM).

Would it be possible to add support for higher-precision quantizations, such as Q4 through Q6 models (e.g., Unsloth's UD-Q4_K_XL to UD-Q6_K_XL for Qwen3.8-Flash-Next: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF)?

Motivation & Context

For users who have workstations or PCs with larger RAM pools (e.g., 96GB–128GB+ RAM) or multi-GPU setups (e.g., dual 3090/4090), running Q4/Q5/Q6 would provide noticeably higher reasoning quality and lower perplexity compared to IQ3/Q2.

Questions:

1. Is supporting Q4+ (or unsloth dynamic/custom quants like UD-Q4_K_XL) on the roadmap?
2. What are the main blockers for higher-bit quant support? (e.g., shard packing format/conversions in ref/tools, ggml tensor type implementations, or CPU-RAM offloading throughput bottlenecks?)
3. Is there a guide or manual conversion workflow to pack and test custom GGUF quants directly with Strata?

Thank you for this incredible project!