Tighter write race checking (#224)

* introduce loopy.Options.insert_additional_gbarriers

* implement WriteRaceChecker

* barrier insertion: use write race checker instead of overlap checker

* tests: gets rid of requirement of certain barriers when there weren't any access races

* should be either l.0.A != l.0.B or l.1.A != l.1.B

* do not intersect the map domains

Intersecting would get ignore some accesses when there's a grid
mismatch => Inaccurate.

* docs

* account for instructions missing hw axes

* "broadcast" statements along the unused hw axes

* fixup! implement WriteRaceChecker

changes related to callables_table

* s/insert_additional_gbarriers/insert_gbarriers/g

* move WriteRaceChecker to loopy.schedule.tools

* extends test_no_barriers_for_non_overlapping_access

* uses markup in _check_for_access_races docs

* describe the special access map values

* do not handroll isl_obj.project_out_except

* adds a test from gh-578