Merge fastcpc

[cr-marcstevens]

• (for speed: forward): progressively increase output set size for the first few steps (t=1,2,3) of forward. this speeds up these first few steps considerably without loss of quality
• (fix forward, backward, connect): make dostep_index volatile, read it at start, write it at end of critical section.
• (improvement for helper): add better support to fix a few bytes at the start of a near-collision block to support more format-restricted usecases for chosen-prefix collisions.
• (for speed: helper): add support to negate a set of diffpaths, in order to negate each 'positive delta m_11'-based precomputed set of upper differential paths instead also computing the 'negate delta m_11'-based set.
• (for speed: helper): add support to combine diffpaths, in order to use precomputed set of upper differential paths and overwrite the full differential paths with the correct ending differences for the near-collision block at hand.
• (fix md5helper): add support to join files consisting of just a differentialpath
• (for scaling & latency: forward): add splitsave parameter to save output over multiple files (for multiple connect processes) in parallel
• (for speed: forward): when full decrease maxcond to stop processing paths that will be pruned later on anyway.
• (for speed: forward): use C++ move instead of copy, as that is faster
• (for scaling: forward): use threadlocal buffers to global container to reduce contention
• (for speed: connect): new mintunnel parameter to prune search from start to avoid bad full paths with too few tunnels, minor code improvements
• (for latency: connect): do parallel read of both inputfiles & new option waitinputfile to wait for each inputfile to exist before trying to load it
• (for scaling: forward, backward, connect): improve input distribution over threads, reducing contention
• (fix avx256 allocation: birthday): older GCCs (like on servers) might not propely do large alignment as required for AVX256
• (for latency: birthday): immediately save birthday collision when found, instead of waiting till all threads finish
• (for scaling: birthday): add saveloadwait parameter that controls how frequently trails are distributed to controllers (saved) and loaded by controllers (load)
• (for scaling: birthday): add generatormode that only generates trails and distributes them to controllers
• (for latency: lib): compressed diffpath archives: configure gzip for best_speed
• (for speed: lib, forward, backward): diffpath fast check solvability: only check changed part
• MacOS fix: differentialpath: clean up operators