Move checkpointing parallelism into TaskExecutor class, use that class for parallel union_by_name

[Mytherin]

#9999 introduced parallel checkpointing by adding a separate mechanism for managing tasks during the checkpointing process. This was necessary because the checkpointing code cannot easily use the normal parallelism loops during execution - as the checkpointing can run outside of regular query execution (e.g. when shutting down a database, or during a commit).

This PR extracts the logic that was added in that PR into a separate class - the TaskExecutor. This class can be used to easily add parallelism in places where we cannot use the regular parallelism model. The TaskExecutor schedules tasks using the TaskScheduler, which are then executed in parallel using the regular worker threads. It merely provides a number of helper functions for keeping track of how many tasks have completed, and for error handling across different threads.

In this PR we use the TaskExecutor to provide a parallel implementation of the union_by_name file scanning. Since we perform auto-detection on all files, this is trivial to parallelize, and can provide substantial speedups when running read_csv or read_parquet with union_by_name enabled over many small files. The union_by_name is also a good showcase for how easy the parallelism is to add using the TaskExecutor, e.g.:

TaskExecutor executor(context);
// schedule tasks for all files
for (auto &file : files) {
	auto task = make_uniq<UnionByReaderTask>(....);
	executor.ScheduleTask(std::move(task));
}
// complete all tasks
executor.WorkOnTasks();

Benchmarks

Below are some timings of reading 1000 small CSV files. Source:

CREATE TABLE t1 AS select i,i,i,i,i,i,i,i,i,i from range(0,2048) tbl(i);
COPY t1 TO 'small_csv.csv' (FORMAT CSV, HEADER 0);
SELECT * from read_csv(repeat(['small_csv.csv'], 1000),delim= ',',  header = 0, union_by_name=true);

v1.0	main	new
3.9s	3.2s	0.6s