bench: Various improvements to trim, substr, split_part benchmarks - #26114

Every string in the substr benchmarks had the same length, and every row
used the same start and count, so every result had the same length too.
The benchmarks also passed a `Utf8View` return field for every input
type and discarded the result, so with debug assertions enabled, the
Utf8 and LargeUtf8 cases returned an internal error that went unnoticed.

Rewrite them like the `left` and `right` benchmarks: draw each input's
length from a range; cover short and long results, results without a
count, and a start computed per row; resolve the return field with
`return_field_from_args`; and check the result. Drop the LargeUtf8 cases,
which run the same code as Utf8, and use a single batch size.

Every string in the split_part benchmarks had the same number of fields,
each of the same length, so every result had the same length. Only two
cases used Utf8View, and both had long results stored out of line.

Rewrite them like the `left` and `right` benchmarks: draw each row's
number of fields and each field's length from a range, and run every
case on both Utf8 and Utf8View. Cover short and long fields, a field
among many, a negative position, a multi-character delimiter, and a
position computed per row, which replaces the cases that passed the
delimiter and position as arrays.

Run each substring_index benchmark on 8192 rows, the batch size the
other string benchmarks use, instead of on 100, 1000, and 10000 rows.