[c++] Integrate SOMAColumn: Arrow adapter methods, part 1

[XanthosXanthopoulos]

This PR replaces the Arrow schema to TileDB schema transformation to use the SOMAColumn create methods.
Also there are a set of new data converters from arrow arrays to std::array for simplification.

This migration also enforces a current domain restriction for string dimensions to libtiledbsoma in addition to the restriction being present only on the R and Python APIs.

[nguyenv]

Can you explain what the Skip and Take are for and document it? It looks like Take is the index of the column to retrieve and Skip is relevant only for geometry columns (where it's always 2)?

Also is there a way to use std::variant or a templated type instead of std::any or would that make things too complicated?

[nguyenv]

Suggested change

	/**
	* Return a copy of the data in a specified column of an arrow table.
	* Complex column types are supported. The for each sub column are an
	* std::array<T, 2> casted as an std::any object.
	*/
	/**
	* Return a copy of the data in a specified column of an arrow table.
	* Complex column types are supported. The type for each sub column is
	* an std::array<T, 2> casted as an std::any object.
	*/

[XanthosXanthopoulos]

Skip and Take are used in 2 places with 2 specific sets on values (either Skip=3 and Take=2 or Skip=0 and Take=2) and are independent of the geometry column. Their usage is to extract specific subranges of ArrowArray data and they come in handy during ArrowSchema -> TileDBSchema where the arrow array provided has 5 values per dimension and we only need the last 2 to set the current domain.

As to using std::variant, adding more SOMAColumn types would require changing multiple variants. The use of std::any here is to enable runtime polymorphism and indirectly introduces a runtime type check (via any_cast, make_any) between the templated function and the actual dimension type. std::variant can provide all the above it is just a different style I am open to discuss further.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.27%. Comparing base (7616147) to head (77a01e1).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3405      +/-   ##
==========================================
+ Coverage   86.22%   86.27%   +0.04%     
==========================================
  Files          55       55              
  Lines        6410     6410              
==========================================
+ Hits         5527     5530       +3     
+ Misses        883      880       -3     

Flag	Coverage Δ
python	86.27% <ø> (+0.04%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Components	Coverage Δ
python_api	86.27% <ø> (+0.04%)	⬆️
libtiledbsoma	∅ <ø> (∅)

[jp-dark]

This doesn't appear to be used anywhere. Can you pull it out into a separate PR and/or add it to the branch where it is used?

[jp-dark]

I didn't review the get_table_column_by_name function, but everything else looks good to me. However, someone with more familiarity with the arrow adapter code should look over this as well.

[johnkerl]

Thanks for working on this @XanthosXanthopoulos !

[johnkerl]

I do seem to recall these were about crash avoidance on some platform (I don't recall which). I'd rather leave these as-is please.

[johnkerl]

This could be pulled out into a utility function

[johnkerl]

Suggested change

	"{} missing",
	"'{}' missing",

Strings in error messages should always be quoted/bracketed. Even if you think it's impossible for the string to ever be empty. Heaven forbid someday there is some bug somewhere somehow ... and an empty string gets in here ... that needs to be clear to everyone that sees the error message.

[johnkerl]

This is a bit awkward.

• column needs to be column_it or some such -- it is not a column, it is an iterator
• Then, const auto column = *column_itafter you check thatcolumn _it != columns.end()`
• Then the rest of these (*column)->foo become column->foo as they should be

[johnkerl]

As on previous PRs: we cannot simply say dom or domain, ever.

There are four things it can mean:

• core domain (which is soma maxdomain)
• soma domain (which is core current domain)

The names are confusing (and too late to change), and confusion is too easy, and developer confusion is high-risk

[johnkerl]

Perhaps rename dom to cdslot

[johnkerl]

Please state as a compact summary, right here as a code comment, what Take and Skip are for, what they do, and an example usage.