run finalizers on their own thread

[JeffBezanson]

In general, finalizers may need to acquire locks, so they should be guaranteed to run on a separate thread to avoid conflicts with the main execution threads. @vtjnash may be able to provide some more details.

[maleadt]

Had to work around that recently, by disabling finalizers when taking a lock that might also be taken during finalizers: JuliaGPU/CuArrays.jl@2258a24

[vtjnash]

Relatedly, I think we should consider renaming jl_gc_enable_finalizers, change the design (#11207), and/or eventually remove this. This function name misleadingly sounds like it prevents finalizers from being run concurrently with a particular region of code. Indeed, the implementation further seems to imply that too. However, with threading, it's not actually true. We could make it true, but I think it's better just to change this to jl_gc_exclude_finalizers_on_thread (after this issue is closed, it'd become a no-op).

[OkonSamuel]

@vtjnash . It would be nice to have the option of preventing finalizers from running concunrently.
That would solve the annoying error in running finalizer: ErrorException("concurrency violation detected") silent warning am getting at MLJ

[mkitti]

Following the test-test-set pattern in the manual helped seems to help resolve an issue we had with HDF5.jl when working with finalizers and locks:
JuliaIO/HDF5.jl#1048

https://docs.julialang.org/en/v1/manual/multi-threading/#Safe-use-of-Finalizers