[IMP] fields: performance of creating with m2m fields

Note: all costs are according to cprofile on my machine, as a fraction
of total time to import, python-flamegraph has somewhat different cost
distributions, YMMV. Profiles use test-import of a 2500 partners file,
with name, country, street and 0~4 tags, all randomly
generated/assigned.

Costs were profiled as split 8.5%/32%/57% between data conversion,
creating records (Model.create/Model._create) and populating stored
computed fields for a total import time (unprofiled) of 1mn (±2s or
something).

For the creation step, 27% (of total runtime) were assigned to the m2m
(fields.Many2many.create), almost all of which (25%) was assigned to
accessing a record's id in a listcomp in a loop.

Optimisations:

* optimise the case where we're only adding / creating values:
  represent the relation as a dict of sets instead of a set of pairs;
  commands 5 and 6 will only operate on the current record's value,
  instead of traversing all pairs in the relation
* optimise accessing the value of an id field by hand-inlining the
  various bits, this is mostly obviated by the other previous
  optimisations (number of Id.__get__ calls reduced from 5.4 millions
  to 0.55) but given how common accessing a record's id is (and
  how it's usually considered to be free) it's still a good idea: cost
  of accessing the id field was reduced by 75% (0.55 million accesses
  went from 3.32% of total runtime to 0.83%)

Without profiling, time to import my test file went from ~1mn to ~45s.

0 files changed, 0 insertions, 0 deletions