[NativeAOT] Implement thin locks

[VSadov]

CoreClr is able to use object header bits for locking in trivial cases. That makes typical use of lock fairly cheap. NativeAOT does not have that and lock(obj) is always a "fat lock".

This is likely the reason for regressions in ConcurrentDictionary benchmarks.
(re: #68891)

[tomcashman]

Where would be a starting point for this? I’d be interested in attempting an implementation to improve concurrency performance on our console ports.

[VSadov]

@tomcashman this is already in progress, sorry.
There is already an implementation at a stage where I know it is correct (as in "passing all tests"), but I do not know yet if the perf is acceptable.

[tomcashman]

No worries I’ll keep an eye out for the PR and we can test performance on consoles 😄