Fix dense [<- writes and support [i, j, k] and selected_ranges()

[mojaveazure]

Fixes the dense [<- write bugs reported in #877, and makes dense writes work for 3D+ arrays.

Bugs fixed

ncol	Before, [i, j]	Before, selected_ranges()	Now
1	silent no-op	silent no-op	written
2	written	error: buffer sizes (writes full domain)	written
3	error: "does not contain all required… columns"	same error	written
4	written	written (only because it fills the domain)	written

• ncol = 1: the single-attribute flatten step required ncol > 1, but its outer if still matched, so the list case was skipped as well. The value kept the name V1, the name check failed, and the if had no else, so the method returned without writing. A name mismatch is now an error.
• ncol = 3: the data.frame-of-dimensions-and-attributes check matched on column count alone (2 dims + 1 attr = 3). For dense arrays it now also requires a matching column name.
• Subarray: was as.integer(c(range(i), range(j))), set only when both i and j were given and there was a single attribute. It was 2D and INT32 only, ignored selected_ranges(), and widened non-contiguous indices. Dense writes with multiple attributes never set a subarray.

New behavior

• arr[i, j, k] <- value on dense arrays, capturing k the same way [ does (nd_index_from_syscall()). A 4th index is an error, not a message as on the read side, because ignoring it would write to the wrong cells. Sparse writes still use i and j only.
• Dense writes honor selected_ranges(), named or unnamed, for any number of dimensions. Dimensions with no selection span their full domain.
• The subarray is typed per dimension (libtiledb_subarray_add_range_with_type()), so INT64 and datetime dimensions work.
• nd_index_from_syscall() no longer evaluates value a second time.

Behavior changes (in NEWS and ?"[<-,tiledb_array-method")

These are now errors:

• dense indices that aren't contiguous and increasing (they used to be widened to their range);
• an index for a dimension the array lacks, e.g. a third index on a 2D array, which used to be silently dropped;
• more than one selected range for a dimension;
• selected_points() on a write;
• a value that doesn't match the array (used to be a silent no-op).

Also: selected_ranges() left on an array object from an earlier read now apply to writes through it.

Known limitation

A dimension with no index or selected range spans its full domain, not its current domain. This only matters for dense arrays that have a current domain set.

Testing

• New inst/tinytest/test_densewrite.R (32 tests):

  • 2D with 1 to 4 columns, for both [i, j] and selected_ranges();
  • an offset block;
  • 3D [i, j, k] and [i, , k], and an array value;
  • named, out-of-order selected_ranges() on a 4D array;
  • 1D, INT64 dimensions, and multiple attributes;
  • every new error.

  It's a separate file because test_tiledbarray.R calls exit_file() partway through on macOS and r-universe.

• Full suite: 1790 tests, 1 failure (test_config_profile.R, rest.token). The same failure occurs on main (1758 tests, 1 failure). This was run against the R changes loaded into an installed 0.34.0 build, not a fresh compile; CI will cover that.

Closes #877

🤖 Generated with Claude Code