Performance regression for full LD matrix

[jeromekelleher]

#1864 updated the LdCalculator code to remove some complicated optimisations. This lead to a performance regression when computing the full LD matrix in some cases (see this comment #1864 (comment))

When developing the new LD APIs (#1900) and algorithms we should compare performance against the code in version 0.3.7 as well as the current main. There were some good ideas in the old code about how to minimise tree transitions using an inner and outer tree when computing lots of pairwise LD values in a predictable pattern, hopefully we can reuse most of this and regain the old performance.

We can close this issue once we have either confirmed we recover the old performance, or decided we don't care.