Re-evaluate values canonicalization in `DictArray::into_canonical`

[lwwmanning]

In #1032 we always canonicalize the dictionary, since that is optimal under the assumption that codes.len() >= values.len(). This is always true in the absence of shared dictionaries.

After we address #252, we may want to take a more nuanced approach. In particular, a #1038 may be a better approach

[lwwmanning]

After #1082 and #1121, the values canonicalization will only really apply to primitive types. Worth testing if we can revert #1032 entirely, possibly alongside some work to speed up BitPackedArray::take (see #1039 )

The benefit of #1032 was by far the greatest on string-heavy datasets (where it called into VarBin::take, which is very expensive, but is obviated by German strings).

Especially on DictArrays with FSST-encoded values, this was kind of pathological prior to #1032:

• we would first do a VarBin take on the encoded FSST strings (copying each encoded string for each instance of its code)
• we would then FSST decompress the now-much-larger pile of compressed strings (full of duplicates)