Fix ArrayArg's update_persistent_hash

See merge request inducer/loopy!332

Fix comparison string for improved conditional printing in test_isl

See merge request inducer/loopy!331

Support zero warm-up rounds in auto_test (Closes #179)

Closes #179

See merge request inducer/loopy!327

Fix deprecated usages in expand_cses()

See merge request inducer/loopy!325

This caused some mysterious warnings in Pytential.

This is a large refactoring, with many pieces:

- Counts from subkernels are incorporated using
  subst_into_{pwqpolynomial,guarded_pwqpolynomial,to_count_map}.
  This replaces a prior, broken scheme that existed on the kernel
  callables branch.

- Separate ToCountMap and ToCountPolynomialMap, i.e. separate
  to-count map types by their value type. The latter type
  now knows (and checks) its isl space.

- The numpy_types argument is now deprecated and ignored, it
  did not seem to do anything previously.

- Introduce Sync() count key for synchronization counting.

- Code/robustness cleanups in the ToCountMap* types.

- All op descriptors now carry a kernel_name.

There are still a few FIMXEs, mainly the SUBGROUP granularity and the
footprint gatherer.

Op/MemAccess: Use .copy() rather than explicit constructor to copy, avoids losing attributes (stats)

Make sure subst_into_pwqpolynomial produces PwQPolynomials that have an output dimension in their space

Fix, test stride mismatch check

See merge request inducer/loopy!323