Fitting normalizer seemingly does all the work on all the gpus

[SrGonao]

When running bergson with a normalizer and multiple gpus, the progress bar constantly jumps back and forth and it seems it estimates normalizers on all the data on all the gpus, instead of splitting it up like it does when computing the index.

[LouisYRYJ]

This should be fixed by feeding

bergson/bergson/build.py

Line 64 in 0f10d84

processor = create_processor(model, ds, cfg, target_modules)

the batches