Switch to a more sensible normalization strategy for loss metrics

[allevitan]

This makes loss metrics a bit more interpretable. I may put some more work into writing better normalizations if I have a similar internet-free train ride in the future, where I can do some cdtools work but don't have a stable connection to any GPUS...

Key notes:

• This affects all reported loss metrics, so after merging this PR, expect the loss metrics for typical reconstructions to change.

Changes:

• Adds a more flexible "normalizer" class to enable different kinds of normalization for the calculated loss metrics when used for reconstructions.
• Removes the intensity-based normalization for the amplitude MSE and switches to reporting a "straight up" amplitude MSE. I have come to believe this is more interpretable in most cases.
• Switches to use a more reasonable normalization and offset correction for the Poisson NLL results, so that they stop being insane values and never go negative anymore. This will hopefully help encourage adoption of the Poisson NLL loss metric.
• Added initializations for all three loss metrics to all reconstruction models.
• Adjusted the tests to account for the new loss values, and ensured that the different losses all see test coverage
• Update the docs to match, including the tutorial section.
• Update the appropriate example script (near_field_ptycho.py, data taken with a photon counting detector) to use the Poisson NLL error metric.