Improvement: reduce global locking in aot_call_function()

[palchikov]

Feature

During execution of aot_call_function() there are up to 3 calls to aot_copy_exception() (for the case of exported function has no or single result, which I believe is the most common case). And each aot_copy_exception() internally locks global mutex _exception_lock. This can be improved by:

1. Remove redundant calls to aot_copy_exception()
2. Make exception_lock() to lock module's mutex instead of global one

Benefit

In my project in a highly concurrent environment I use hundreds of instantiated modules, and each module processes about 1k requests per second. While each module performance is limited by execution time of exported function, I need to increase the number of modules to gain sufficient throughput. With global locks I have very strict limits to which I can add new executors. The more modules I add, the slower each of them works.
By removing global locks (or at least by reducing the number of locks required to aot_call_function()) we can achieve linear scalability.

Implementation

1. Reducing the number of aot_copy_exception(). There are 3 calls I mentioned above: inside invoke_native_internal, then one for debugging purpose, and finally just before return. invoke_native_internal() uses exception check as its whole result, so further calls to aot_copy_exception() seems to be redundant. The branch functions with result_count > 1 reuses return result in similar places.
2. The comment inside exception_lock() body tells that there were some plans to make a mutex belong to a particular module. Probably this requires complex changes, but it will give great performance boost in highly concurrent environments.

[lum1n0us]

#4059 has been merged, and we are planning to close this issue. Please assist us by confirming.