[Clang] Fix cluster_dims thread block limit checks (#230186)
The limit is on the total number of thread blocks, 8 for NVPTX and 16
for AMDGPU, but two checks got in the way. A 4-bit cap on each dimension
rejected cluster_dims(16, 1, 1) while accepting the same 16 blocks as
cluster_dims(4, 4, 1), and an omitted dimension contributed 0 rather
than 1, zeroing the product so cluster_dims(15, 15) compiled cleanly.
Bound each dimension only enough to keep the product from overflowing
and let the existing total size diagnostic enforce the limit. Code that
relied on the unenforced limit now fails to compile.