Add ConstrainedPartitionsIntRange

[gsakkis]

This PR introduces ConstrainedPartitionsIntRange, an IntRange subclass that will replace the latter in DenseTensorSchema in a follow-up PR.

A ConstrainedPartitionsIntRange is an IntRange whose every subrange (with the possible exception of the first) generated by partition_by_count or partition_by_weight starts from a given range of offsets (start_offsets), in contrast to a regular IntRange whose subranges can start from any offset within the original range.

Example:

In [2]: ir = IntRange(10, 40)

In [3]: [(r.min, r.max) for r in ir.partition_by_count(3)]
Out[3]: [(10, 20), (21, 30), (31, 40)]

In [4]: cpir = ConstrainedPartitionsIntRange(10, 40, start_offsets=range(0, 100, 4))

In [5]: [(r.min, r.max) for r in cpir.partition_by_count(3)]
Out[5]: [(10, 19), (20, 31), (32, 40)]

In this example every subrange (with the possible exception of the first) of cpir starts with a member of start_offsets=range(0, 100, 4).

While at the ranges module, two unrelated changes are included:

• Fix InclusiveRange.equal_values() to compare correctly IntRange with WeightedRange instances.
• Drop the InclusiveRange.factory staticmethod; not worth the complexity.

[georgeSkoumas]

LGTM!

[georgeSkoumas]

@gsakkis do we need all the checks here or np.all(self.values == other.values) is enough?

[gsakkis]

At minimal we also need len(self) == len(other) because comparing arrays of different size raises DeprecationWarning (and will raise an error in the future). The min/max comparisons are not strictly necessary, they're just an optimization (in case they are false).