I leveraged GitHub actions to measure optimization results on various systems. Only differences between bytearray and bytes, memoryview are significant, since benchmarks might not be run on the identical hardware.
python -m timeit -s "a=bytearray(4096);b=b'x'*1024;" "a[:1024]=b"
python -m timeit -s "a=bytearray(4096);b=memoryview(b'x'*1024);" "a[:1024]=b"
python -m timeit -s "a=bytearray(4096);b=bytearray(b'x'*1024);" "a[:1024]=b"

Before: https://github.com/msoxzw/cpython/actions/runs/4120204201

After: https://github.com/msoxzw/cpython/actions/runs/4120296060

Therefore, this PR makes bytearray slice assignment for bytes-like object run as fast as that for bytearray.

I manage to benchmark only on Windows on the same machine through GitHub actions. Buffer protocol would incur marginal ~10% overhead, if bytearray is regarded as buffer object.
https://github.com/msoxzw/cpython/actions/runs/4238205216/jobs/7365033501

If data copies are bypassed only for bytes objects, such performance penalty would be avoided accordingly.
https://github.com/msoxzw/cpython/actions/runs/4238204896/jobs/7365028180

So is it acceptable to such performance overhead?

How does it work for ba[:] = memoryview(b'abcd')[::2]?

Thanks for such wonderful review.

It would work like byte or bytearray concatenation: bytearray() + memoryview(b'abcd')[::2]

This behavior is dictated by PyBUF_SIMPLE flag in PyObject_GetBuffer function. Nevertheless, it is probable to preserve original behavior, and thus maintain API compatibility.