Transform coordinates by a RateMap

[nspope]

Some infrastructure for supporting mutation rate maps

[nspope]

please take a glance at this when you have a chance @hyanwong . The plan is ultimately to internally transform the tree sequence to a "mutational unit" coordinate system (which might imply removing some edges/mutations in zero-rate intervals), do the dating, then copy the dates/posteriors over the the tree sequence in the original coordinate system. There's some tricky details to sort out, so I'm going to do this in pieces.

[hyanwong]

LGTM. Two very minor comments.

[hyanwong]

This is a great start to being able to test a more flexible definition of mutation rates. I see that the function can be used to mask out a region by setting the underlying rate to NaN, but might we eventually also want a more standard masking approach, e.g. which doesn't change the underlying coordinate system?

I think this is fine for the time being, though.

[nspope]

Thanks!

but might we eventually also want a more standard masking approach, e.g. which doesn't change the underlying coordinate system?

Not sure I'm understanding what you mean-- we won't be changing the coordinate system in the output. The coordinate system will only be changed internally for the EP algorithm, to express things in terms of mutational units, which lets us support ratemaps with minimal changes to the code. For example, even if the mutation rate is constant, unsequenced intervals could be modeled as zero-rate intervals: this is similar to using delete_intervals but doesn't "chop up" edges that contain small gaps.

There's already tskit infrastructure for "hard masking" via delete_intervals, so I don't think we need to do anything there.

[hyanwong]

Thanks!

but might we eventually also want a more standard masking approach, e.g. which doesn't change the underlying coordinate system?

Not sure I'm understanding what you mean-- we won't be changing the coordinate system in the output. The coordinate system will only be changed internally for the EP algorithm, to express things in terms of mutational units, which lets us support ratemaps with minimal changes to the code. For example, even if the mutation rate is constant, unsequenced intervals could be modeled as zero-rate intervals: this is similar to using delete_intervals but doesn't "chop up" edges that contain small gaps.

There's already tskit infrastructure for "hard masking" via delete_intervals, so I don't think we need to do anything there.

Ah, this is a good plan, internally. Thanks for merging. Also I see you suggested this as an internal approach for tskit "soft masking", which seems like a good start.