to generically implement upgrade/downgrade operations. If scoped_lock_type
implements upgrade_to_writer and downgrade_to_reader then traits class
will just call those methods, otherwise it will provide default implementation.

Hashmap code cleanup

concurrent_hash_map: introduce version field and add reserved space

Fix array move assignment operator

to make sure that data was create using the same (or compatible) version
of concurrent_hash_map.

Also change order of some members so that embedded_segment (which size is
dependent on size of the mutexes used is at the end). This allows to have
segment_enable_mutex aligned to cacheline.

* Expect bucket bucket to be already locked inside get_node().
* Move logic for restarting search after mask race to internal_find/insert.

explaining that it does not add moved-from array to the transaction.
Also add test for move ctor.

Until now, move assignment for array performed copy instead of move.
Fix this issue and add test for move assignemt.

Update doxygen files

This function should only be called after modification (only in internal_insert).

concurrent_hash_map: pass reference to bucket instead of pointer in get_node

Copied from pmem/pmdk.

Using uint64_t removes padding.

get_node() expects pointer to bucket_accessor to always be set, so it's safer to
use reference.

concurrent_hash_map: remove wrong assert in get_node method

Template locks

Change lock implementation in hashmap to use templates instead of
compile time parameter.

Changes are related to pretty complicated c++ syntax for
template derived class using a member it inherits from its template base class.
Coding convention is:
* To get types defined in base class in derived class
        using name_of_type = typename base_class::name_of_type
* To get access to methods defined in base class
        using base_class:method_name
* To get access to fields defined in base class
        this->field_name

As hash_map_node_base is now template, it's impossible to determine type
of empty_bucket defined in way it was until now - so now it's removed.
IMHO checking if node is nullptr is cleaner than comparing it to
empty_bucket(which is nullptr anyway)

Those functions are redundant, and contain no more needed
compartions pointer to 63. Aplication logic without such calls seems to
be little bit more readable

Condition in the assert is not guaranteed to be true - it was
added by mistake in: 7d7eebac

It's needed to be used by user as template parameter.

 pmem::obj::experimental::internal::shared_mutex_scoped_lock ->
 pmem::obj::experimental::shared_mutex_scoped_lock

This change may be not displayed properly by git.
Try:
git show SHA --anchored="namespace"

Fix string sso usage

Use non_sso_data() and sso_data() methods instead of direct access to
make sure data is accessed only when it is initialized.

In insert, it was wrongly assumed that if size < sso_capacity it means
that sso is used. It is possible that non_sso type is used and
size is < sso_capacity so explicit check is needed.

concurrent_hash_map: reimplement my_mask

Fix release build errors

my_mask is a variable which is reinitialized on each process
restart. Until now, v<> was used for this purpose. Because v<> is not
productized yet, reinitialize my_mask in runtime_initialize() method. Now,
runtime_initialize() MUST be called after proccess restart.

In future, when v<> is productized, we may switch back to using it.

and call it in every test before using hashmap

concurrent_hash_map: implement insert using tx

concurrent_hash_map: add test for concurrent inserts with the same key

Hash map minor refactoring

travis: limit the coverage build

Implement inserts which take rvalue ref uisng internal_insert.
This is more consistent with implementation of other inserts.