Translate weights in sum mode

[thehrh]

Motivation

Meant to complete what #934 began: probability weights shouldn't transform in "average" mode by default, but in "sum" mode.

• this implies that under a transformation to a histogram/map representation, the Container class will sum—and not average—weights
• the inverse ("unbinning") or changing the binning ("rebinning") is not generally possible without making approximations and is thus not implemented (this is currently the case in PISA and isn't changed by this PR)
• in "average" mode, however, PISA performs a) a simple bin-value lookup for unbinning and b) resampling for rebinning, which is problematic for our probability weights due to the resulting change in normalisation (verify this by transforming a weight event-count histogram (back) to "events" representation or by merging bins)

Changes

• add "weights" and "initial_weights" keys to so-far empty Container.sum_mode_keys attribute
• change hierarchy of if statements in Container.translate(), checking for translation route—not mode— first, and
  • no longer distinguish between "average" and "sum" modes in case of one-to-one relationship, i.e., for "events" ↔ "log_events" (distinction only relevant when aggregating/unbinning/rebinning; NB: before, for undocumented reasons, these routes were rejected for "sum" mode)
• additionally, supply method Container.set_item_no_invalidate(), which doesn't invalidate any representations
• expand test_container() unit tests to demonstrate expected behaviour and ensure sanity
• adapt test_Pipeline() unit tests because "weights" cannot be translated from binned to "events" output any longer
• major rewrite of utils.hist's setup_function for the (unconventional) case of a binned calc_mode
  • TODO
• slightly expand and fix pisa_modes.ipynb introductory text on representations/modes and gently restructure cells to execute
• expand logging and documentation
• also fix a spelling mistake in run_unit_tests.py