Improve allocation optimization with gc_preserve intrinsics

[yuyichao]

Delete them directly so that they don't mess up LLVM optimizations.

[maleadt]

I think this broke (or at least changed behavior) of multi-arg @gc_preserve as used in CUDAdrv:

function foobar(a, b)
    a_ref = Base.RefValue(a)
    b_ref = Base.RefValue(b)

    a_ptr = Base.unsafe_convert(Ptr{Void}, a_ref)
    b_ptr = Base.unsafe_convert(Ptr{Void}, b_ref)

    ptrs = [a_ptr, b_ptr]
    Base.@gc_preserve a_ref b_ref ptrs

    ccall(:whatever, Cint, (Ptr{Ptr{Void}},), ptrs)
end

const a = [1]
const b = [2]

foobar(pointer(a), pointer(b))
@code_llvm foobar(pointer(a), pointer(b))

Before merging this, both lifetime.ends are together near the end of the function, while now one lifetime is ended much sooner. The old behavior can be restored by using one @gc_preserve for each argument (a_ref and b_ref). Or am I misusing the intrinsic?

[yuyichao]

#24475