Make master outlive confdata out of shared memory errors

[DrDet]

Here're two new approaches:

Events throttling

To prevent OOM errors on confdata overflows we start to throttle events, when too little shared memory left. There're two modes:

• Soft OOM (85%) - replay only existing keys related events until server restart. Ignore new keys.
• Hard OOM (95%) - replay nothing. Ignore all events, don't read binlog at all.

If server enters any OOM modes, it can't go back out of it. Manual server restart is needed for that.

External preventive allocation heuristic checks

Before any action, that internally invokes allocations, we explicitly check that allocator has enough memory for that.
To achieve this we need to somehow calculate upper-bound for memory that will be required by the action in each particular case.
There're two 'kinds' of allocations:

1. Allocation made by our runtime (array<mixed>::mutate_if_shared(), etc..). It's not too complicated, because it's our own code and we can calculate it easily.
2. Allocation made by STL, out of std::map<> which uses our shared memory allocator. To handle this we will use common sense, some knowledge about search trees internal structure and compile time upper-bound for tree node size from our allocator:

   kphp/runtime/memory_resource/resource_allocator.h

   Line 48 in d685b6c

   static constexpr size_t max_value_type_size() {

New stats for confdata

This PR also includes writing confdata metrics to statshouse:

• kphp_confdata_memory
• kphp_confdata_events
• kphp_confdata_sections_by_count
• kphp_confdata_update_fails

New option for blacklisting confdata events

--confdata-force-ignore-keys-prefix — an emergency option, e.g. 'highload.vid*', to forcibly drop keys from snapshot/binlog; may be used multiple times.

Timeouts for reading binlog

To prevent master hanging in master_cron(), a timeout for reading a next binlog chunk is added, controlled by --confdata-update-timeout. Every second (every cron iteration) we allow binlog reader to consume next event no more than timeout seconds. That allows master to perform other work.

Note, that an option for reading binlog in a separate thread can't be implemented, since the master process is also fundamentally single-threaded (lots of globals, global allocator, global allocator switches between heap/shared memory, etc.).