Iterator operator++/--:
 Removing nodes/leaves which are pointed to be iterators might
 result in operator++/operator-- failures (since nodes/leaves are
 detached from the tree find_leaf/next_leaf functions will not
 find a proper entry).

 Fix this by adding a fallback path which just searches from the top
 (using lower_bound/upper_bound).

internal_bound:
 Do not traverse the tree twice. Instead save the path from first
 traversal to a vector and just iterate over the vector. Before
 returning iterator to the user, we can check if the path changed
 (if any node on the path was modified). If so, the operation needs
 to be retried.

Implement defer free for radix_tree

and enable overwrite tests (since runtime_initialize_mt is
available).

which enables single-writer multi-reader concurrency. After calling
runtime_initialize_mt() each assignment (done through assing_val or
insert_or_assign) does not modify the value in-place but rather
allocates new node. Also, erase does not free nodes/leaves immediately
but adds them to a garbage list which can be processed using
garbage_collect.

gha: temporarily disable codecov

gha: temporarily disable windows builds

Update dockers

and clean up images a little.

Various fixes

radix_tree: use std::atomic<tagged_pointer>

add readme to benchmarks and fix dividing by zero in radix bench

tests: fix sign-conversion warning in helpers_test

add benchmark which compares radix to std::map

Make syncthreads() in parallel_xexec multi-use

Previously syncthreads could only be used once and this was not
documented. This patch allows for having multiple synchronization
points in a function. This might useful for testing hazard pointers.

Add example for using pool as a class member

CMake fixes

version comparison should be used, instead of string comparison.
Simplify the function, by the way.

- extend print messages,
- print warnings once per build, where possible,
- add more comments,
- fix minor todo's,
- add cache string for TBB_DIR.

Clean docker images