[dzannotti]

Stacked on #34 because it needs the GGUF imatrix statistics from upstream ggml-org#14891, which #34 brings in. Once #34 is merged and its branch deleted, GitHub retargets this PR to master.

Port of ggml-org#15550 (EAddario, still open upstream) at 839a43500. It adds --target-bpw, --target-size, --state-file and --target-exclude to llama-quantize. For each tensor it picks the quant type that reaches the overall bpw or file-size target at the lowest estimated error, using the activation statistics that ggml-org#14891 writes into GGUF imatrices.

• Applied as the PR's diff against its upstream base cb7934c52, as one commit. The PR's stray 1k-line uv.lock is left out.
• common/imatrix-loader.*: the PR's .stats / stats_schema loading is combined with imatrix: calculate activation-based statistics for new format (GGUF) imatrices ggml-org/llama.cpp#14891's n_layer_nextn. Schema names and the .stats tensor suffix match what the merged llama-imatrix writes.
• When upstream merges quantize: add option to automatically choose optimal quant types to reach a file/bpw target size at lowest error ggml-org/llama.cpp#15550 (possibly changed), the next sync will conflict in llama-quant.cpp / quantize.cpp. Take upstream's side there.

Measured

Qwen3-1.7B from Q8_0 (--allow-requantize). The imatrix comes from docs/ (60×512 tokens). Perplexity is on held-out tools/ + examples/ READMEs (60×512 tokens), Vulkan:

model	size	BPW	PPL
Q8_0 (source)	2059 MiB	8.50	6.4084 ± 0.166
Q4_K_M + imatrix	1217 MiB	5.03	6.4925 ± 0.169
--target-bpw 5.0262	1217 MiB	5.03	6.4753 ± 0.168
--target-bpw 4.9	1187 MiB	4.90	6.5327 ± 0.170
--target-size 900m	894 MiB	3.69	6.8358 ± 0.179

Targets are hit precisely. At Q4_K_M's exact size the chosen mix loses about 20% less perplexity versus Q8_0 than Q4_K_M does. That is inside the error bars on this small model and calibration set, so treat it as encouraging rather than proven. Reusing --state-file brings a re-target down from minutes to 11 s.