feat(gguf-tools, cli): add --strip-nextn for community fine-tunes and prompt support in /read

[Trecto34]

Summary

This PR introduces two enhancements to improve compatibility with community fine-tunes and interactive vision usability without affecting default model paths:

1. gguf-tools/qwen36-quantize (--strip-nextn / --strip-mtp):

   • Community fine-tunes and abliterated models (e.g. Huihui-Qwen3.6-35B-A3B-abliterated, RavenX, etc.) often include a 41st speculative MTP layer (blk.40.*), setting block_count = 41 and nextn_predict_layers = 1.
   • Running these directly causes runtime layer and dimension mismatches in Q36.
   • Adding --strip-nextn:
     • Strips blk.40.* and .nextn_ tensors during parsing.
     • Strips nextn_predict_layers KV metadata record.
     • Safely patches GGUF block_count from 41 to 40 via strict key and type checking (ends(key, ".block_count") && type == GGUF_TYPE_UINT32 && val == 41).
     • Non-breaking: Default quantization path remains completely untouched when the flag is omitted.

2. q36_cli (/read FILE [PROMPT]):

   • Allows passing an optional user prompt alongside image files in interactive CLI mode:
     • /read photo.jpg Describe what is in this image
     • /read "path with spaces/image.png" What text is written here?
   • Supports quoted paths (single/double quotes) containing spaces.
   • Synchronizes usage and help strings across q36_cli.c and q36_help.c.

3. Documentation:

   • Updated README.md and gguf-tools/README.md with instructions for quantizing and running community fine-tunes and abliterated models.

---

Benchmarks & Verification

Tested on AMD BC-250 APU (RADV GFX1013 Vulkan, 15.35 GiB UMA).

1. Build Verification

All targets built cleanly with 0 errors and 0 warnings:

make clean
make vulkan-generic
make vulkan-bc250
make cpu
make -C gguf-tools

2. Correctness Regression Tests

./q36_test --server
# Output:
# server: OK
# q36 tests: ok

3. Throughput Benchmark Comparison (q36-bench)

Smoke benchmark command:

./q36-bench -m <model> --prompt-file tests/long_context_story_prompt.txt --ctx-start 32 --ctx-max 64 --ctx-alloc 128 --gen-tokens 8

Default Model (q36moe.gguf - official Qwen3.6-35B-A3B IQ2_XXS/Q2_K):

ctx_tokens,prefill_tokens,prefill_tps,gen_tokens,gen_tps,kvcache_bytes
32,32,152.96,8,91.63,140
64,32,228.12,8,92.50,268

Abliterated Community Model (Huihui-Qwen3.6-35B-A3B-Abliterated-Q36-IQ2XXS.gguf quantized with --strip-nextn):

ctx_tokens,prefill_tokens,prefill_tps,gen_tokens,gen_tps,kvcache_bytes
32,32,125.12,8,91.71,140
64,32,227.46,8,91.20,268

• Generation/decode throughput is virtually identical (~91.6 vs ~91.7 t/s).
• Prefill scaling matches the default model (~228 vs ~227 t/s at 64 tokens).