Be able to keep intermediate imatrix results

[ikawrakow]

It seems there is a lot of experimentation with the importance matrix going on, and user would like to be able to compute imatrix results for different numbers of tokens in one go.

This PR adds this ability:

./imatrix -m <some_model> -o <some_file> --keep-imatrix 50 [other parameters]

will result in imatrix results being kept for each number of chunks divisible by 50 in some_file.at_50, some_file.at_100, etc.

[Nexesenex]

Great job!
And because I'm abusive, here's another idea.

As soon as the matrix calculation is not fully offloaded on GPU, calculating thousands of chunks straight ain't short, especially for big models.

Would it be possible to resume iMatrix calculations from a previous checkpoint created by this PR at a given ctx size, considering that (and I'm captain obvious here) at equal ctx size the calculations seem to be identical from one run to another at each given step?

Ideally, there should be a warning if the user selects a matrix file with a different ctx size to resume a Matrix calculation (it's easy to mess batch files with ctx numbers ^^). But not for the model itself, for renaming and fixing typos in model naming is quite common, at least for me!