Reduce memory usage and speed things up when shimming

[po09i]

Description

When shimming with large coil profiles and/or many target slices, we can end up using a large amount of memory (>10Gb). I have isolated the problem to be when we calculate the shimmed profiles. This causes Shimming toolbox to be really slow since it saturates the RAM.

This PR:

• Removes variables that consumed lots of memory but were not used afterwards.
• Adds a decorator to time a function and output it in debug logger mode
• Improves the stability of the optimizer by scaling the output of the _residual function

[po09i]

Current improvements using coil profiles (128x128x75), fieldmap (127x127x55), target (110x110x55). Here is the peak RAM usage.

Before: 14.5GB

After: 4.13GB

Total time to process 156s:

• Sequencer: 134s
  • processing: 112s
  • Calculating theoretical shim: 22s
• Plot coefs: 20s
• Other: 2s

[gaspardcereza]

LGTM. The timeit decorator idea is great !