Consider maintaining separate Module `block_dim` variables for CPU and GPU

[shi-eric]

With the addition of a CPU fallback for the tile API in 83a5845, we can get into situations where modules are compiled on the GPU using a nonsense block_dim of 1, leading to unnecessary module recompilation when a GPU kernel is launched with the intended number of threads per block.

[c0d1f1ed]

This is affecting the APIC tests added as part of #1431. The recompile calls cuModuleLoadDataEx during CUDA graph capture, which is unsupported on drivers < 12.3. I have a fix in the works.