[Clang][OpenMP] Insert alloca for kernel args at function entry block instead of the launch point.

If an inlined kernel is called in a loop, the launch point alloca would
lead to increasing stack usage every time the kernel is invoked. This
could make the application run out of stack space and crash. This problem
is fixed by using the alloca insertion point while creating the alloca instruction.

Fixes https://github.com/llvm/llvm-project/issues/60602

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D145820