[llvm] [AMDGPU] Don't DEALLOC_VGPRS from callable functions (PR #72245)

[Jannik Silvanus via llvm-commits]

jasilvanus wrote:

Wasn't there also the issue that we shouldn't send the message if the current function contains calls, because a callee might have accessed scratch? Is this handled elsewhere?

https://github.com/llvm/llvm-project/pull/72245