improve the rwalk behaviour for difficult cases.

[segasai]

This is a small but useful improvement. When doing random walk sampling the walking consists of two parts.

• Generating a point in a random ball/ellipse around the current point.
• Testing if it is accepted based on logl criteria

These steps are repeated 'walks' times. Sometimes even the first step generation of a point can fail, because say the ellipsoid is longer than the cube and a generated point is outside. Now I do not consider that as a 'walk'. I only consider walk attempts that are within the cube and that get a function evaluation. That should make things more stable and reduce situations when the none of walking steps are accepted (i.e. #354 )

[coveralls]

Pull Request Test Coverage Report for Build 2870843094

• 3 of 4 (75.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.02%) to 90.414%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
py/dynesty/sampling.py	3	4	75.0%
<!--	Total:	3	4	75.0%	-->

Totals
Change from base Build 2863436648:	-0.02%
Covered Lines:	3669
Relevant Lines:	4058

---

💛 - Coveralls

[segasai]

That was not a good idea. that still breaks the detailed balance. Closing.