Parallel mark performance degradation caused by incremental collection

[ivmai]

I see GC is slower now with parallel_mark+incremental vs parallel_mark.

Source: master (6dcc828)
Host: Ubuntu 22.04 / x86_64
Compiler: gcc 11.4.0
How to build, e.g.: gcc -I include/ -I libatomic_ops/src -D GC_THREADS -D THREAD_LOCAL_ALLOC -D PARALLEL_MARK -O3 -march=native -Wall -Wextra -Wpedantic -D USE_MMAP -D USE_MUNMAP -o gctest tests/gctest.c extra/gc.c -lpthread -ldl

Output:

> ./gctest
Running on LINUX/X86_64 target
Supported VDBs: manual soft mprotect
Switched to incremental mode
Reading dirty bits from /proc
Completed 6 tests (concurrently)
...
Full collections took 309 ms
World-stopped pauses took 3772 ms (1514 us each in avg.)
Completed 3085 collections (using 16 marker threads)
Collector appears to work (testing done in 7231 ms)

> GC_DISABLE_INCREMENTAL=1 ./gctest
Running on LINUX/X86_64 target
Supported VDBs: manual soft mprotect
Completed 6 tests (concurrently)
...
Full collections took 1203 ms
World-stopped pauses took 1042 ms (299 us each in avg.)
Completed 1483 collections (using 16 marker threads)
Collector appears to work (testing done in 2305 ms)

Related: #151

[ivmai]

Also related: #545