Refactor/fixes

[timcadman]

Background

Fixes found in a full review of the v7.0 refactor (batches 1–10): a data leak, behaviour that changed relative to 6.3.5, and small doc and code errors.

What's changed

• ds.asList data leak: asListDS returned the converted object to the client. It is now an assign function. It was an aggregate only to work around an old repository quirk with lists; tests confirm lists now assign correctly on Opal and Armadillo.
• NULL objects: .loadServersideObject now errors on a NULL object ("is NULL") for every function, including ds.class, which used to return "NULL". Deliberate breaking change: NULL objects only arise from mistakes such as a typo in ds.make.
• Class rules restored to 6.3.5 behaviour:
  • meanDS/varDS accept logical, so an all-NA column no longer breaks the call.
  • corDS/covDS accept factors again (correlated on level codes).
  • glm/glmSLMA/glmer/lmer/gamlss accept data.frame or list, not matrix, matching what the model functions accept.
  • lengthDS/dimDS accept any class.
  • meanSdGpDS accepts a numeric grouping variable.
• Behaviour restored:
  • ds.cor/ds.cov ignore y (with a warning) when x is a matrix or data.frame.
  • ds.changeRefGroup treats a numeric ref as a level name again.
  • BooleDS accepts "TRUE"/"FALSE"/"NA" as V2.
  • levelsDS checks the privacy level first.
• ds.mdPattern combine: pools columns by variable name, not position. md.pattern orders columns by missingness, which differs between studies, so variables could be mixed.
• seqDS: from/to/by/length.out are converted with as.numeric instead of being looked up as object names.
• glmChecks, ds.table: restore the user's warn option on exit. This fix was on 6.3.x but never ported to v7.