[llvm] [Offload] Allow CUDA Kernels to use arbitrarily large shared memory (PR #145963)

[Giorgi Gvalia via llvm-commits]

gvalson wrote:

> Based on CUDA document, the former can be set by `cuFuncSetAttribute`. If `MaxDynCGroupMemLimit` corresponds to the former, it needs to be mutable; otherwise it should be set to the latter value (and then immutable).

It does correspond to `CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES`.

https://github.com/llvm/llvm-project/pull/145963