Move duplicated imatrix code into single common imatrix-loader.cpp

[bartowski1182]

Overview

quantize.cpp and imatrix.cpp duplicated the same code for loading the imatrix

This change pulls those functions out to a common file with the same imatrix and legacy imatrix loading functions

Each handled post-loading differently with imatrix.cpp keeping the raw sums and counts so they can be merged (with --in-file), whereas quantize.cpp normalized immediately, so the new common_imatrix struct stores the information as it's loaded and each can do with it as they need

Also fixed a minor bug that caused --show-statistics to fail (params weren't set before being used) so that I could use it to compare before and after

Additional information

Tested parity of before and after by running the following scenarios and running sha256sum on the results:

./build/bin/llama-imatrix --in-file ../google_gemma-4-E4B-it-imatrix.gguf -m ../google_gemma-4-E4B-it-bf16.gguf -f ../groups_merged.txt -o test.gguf
./build/bin/llama-imatrix -m ../google_gemma-4-E4B-it-bf16.gguf -f ../groups_merged.txt -o test-legacy.dat --output-format dat
./build/bin/llama-imatrix --in-file test-legacy.dat -m ../google_gemma-4-E4B-it-bf16.gguf -f ../groups_merged.txt -o test-legacy-in.dat--output-format dat
./build/bin/llama-quantize --imatrix test.gguf ../google_gemma-4-E4B-it-bf16.gguf ./gemma-4-test-Q4_K_M.gguf Q4_K_M
./build/bin/llama-quantize --imatrix test-legacy.dat ../google_gemma-4-E4B-it-bf16.gguf ./gemma-4-test-Q4_K_M.gguf Q4_K_M

I also ran --show-statistics since that's the other place the loader is used with imatrix.cpp and the results are the same before and after as well

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, primary method of deduplication, all testing ran manually

(ironically a duplicated PR because I'm bad at git)

[bartowski1182]

@ggerganov can I get another review on this one?