Command that calculates some statistics over the errors introduced by
quantization, at the moment mean square error and max error for layer
weights. Should be useful for testing quantization improvements.

Needs some internal state from ggml and llama that should not be part of
the public API.

Show some error percentiles, should be less noisy than just the max error.

Test quantization in smaller chunks instead of layer-at-a-time.

Show RMSE instead of MSE - keeps similar range to the other metrics.
Regex match on layer pattern.

Expose reference quantization implementation and add option to use it
for tests.

Move into main header with comment not to use, per PR feedback