If the tree contains only a single element smaller than the
looked-for key, radix_tree tried to call make_iterator on nullptr.

This resulted in segfaults in pmemkv tests.

Bug found with the help of UBSAN.

Merge stable-1.12 into master

Mpsc queue remove overload with copying data in lambda function

Merge stable-1.11 into stable-1.12

It seems it is unecessary compicated to use.

Merge stable-1.10 into stable-1.11

Merge stable-1.9 into stable-1.10

Fix mpsc queue const

CI: remove Travis CI configuration on stable branches

Set secret github token, for auto doc update.

Merge stable-1.8 into stable-1.9

tests: link atomic to all tests

Not only self_relative_ptr tests use atomics. On some platforms
(e.g. riscv64) -latomic is needed for a few other tests
(pool_cleanup, volatile_state).

Merge stable-1.7 into stable-1.8

Merge stable-1.6 into stable-1.7

[stable-1.9] docker: fix all images

- fix missing dependencies,
- fix images' FROM path,
- bump dependencies' versions to omit known issues.

tests: do not use basic_string_view comparison as a workaround

Radix tree: make mt mode a template parameter

tests: add libunwind suppresion

mpsc_queue: do not call ringbuf_consume/release with number of byte

Memcheck reports an error in libunwind exception raising.

for C2593: 'operator ==' is ambigous on Visual Studio 14.0

if !MtMode. This improves performance for single-threaded workloads.

After this change performance difference when using MtMode true vs false
is ~ 15% for readseq in pmemkv-bench benchmarks for radix.

This allows for better optimization and getting rid of atomic
operations in single-threaded workloads.

but rather with number of cachelines. This allows to workaround
the size limitation.