review simplify=TRUE rules in frollapply()

[jangorecki]

After #7314 our simplify rules got very wide. The PR meant to add handling for adaptive rolling median.

Considering frollapply manual says

In future version we may improve internal \code{simplifylist} function, then \code{simplify=TRUE} may return object of a different type, breaking downstream code. If results are already simplified with \code{simplify=TRUE}, then it can be considered backward compatible.

it is risky to keep current simplify rules and this statement together.

We should either:

1. change manual that even simplified results may not be considered backward compatible
2. or make simplify to work for a less cases

If we go with 2. then this issue can be good for agreeing on the scope of simplifylist().

1. absolute minimum are:

• coerce int to double when varies within answer results (frollapply simplifylist could be more smart about median results #7313)
  list(NA, 1.5, 2L) -> c(NA, 1.5, 2)
• coerce fill to answers type
  list(NA, 1L, 2L) -> c(NA_integer_, 1L, 2L)
list(-9, 1L, 2L) -> c(-9L, 1L, 2L)
• unlist for atomic vector of matching types
  list(1, 2, 3) -> c(1,2,3)
• rbindlist of list/df/dt of matching dimension and types
  list(dt(a=1, b=2), dt(a=2, b=3)) - > dt(a=1:2, b=2:3)

2. nice to have

• recycle fill to answers length
  list(NA, 1:2) -> list(c(NA,NA), 1:2)
• match fill names to answers names
  list(NA, c(a=1,b=2)) -> list(c(a=NA,b=NA), c(a=1,b=2))
• rbindlist(lapply(x, as.list)) of length 2+ vectors
  list(1:2, 2:3) - > dt(1:2, 2:3)