If you don't need "predictable randomness", like for repeatable statistical simulations, then absolutely, you should only use getrandom(). On recent Linux, this is implemented in the vDSO and is super fast. Few excuses now to use anything different.
The portable API is getentropy, which glibc provides as a simple wrapper around getrandom. getentropy was added to POSIX, and is also available on most modern unix systems, including FreeBSD, Illumos, NetBSD, macOS, OpenBSD, and Solaris.
arc4random has been provided by glibc 2.36 (2022), and is available on all the above-mentioned systems as well. If you don't want to make a syscall per request (outside Linux), just use arc4random; it'll be the fastest method available. musl libc lacks arc4random, unfortunately, but you can always ship a small wrapper.
Systems that support arc4random also support arc4random_uniform, which is a way to get an unbiased unsigned integer between 0 and N (up to 2^32-1). That's probably the most important reason to use the arc4random family.
vDSO getrandom has been in the kernel for what, two weeks? And it is only "super fast" compared to the unbelievably slow full syscall. Compared to PCG it is like watching rocks grow.
Linus merged it on July 24, 2024, so about a year I guess. Kernel is released ~8 weeks after the merge window, so I suppose September or so.
I think neither are unbelievably slow. I dunno, take some measurements and see, maybe it suits you.