Fix divisibility computation exceeding max_divisor. (#153)

`compute_divisibility` used `<=` instead of `<` in its loop bound,
allowing the result to overshoot `max_divisor`. With the default
max_divisor=16, any shape divisible by 32 (common in GPU workloads)
would produce a `DivBy(32)` assume in the Tile IR bytecode instead of
the intended `DivBy(16)`, mismatching cuTile Python's behavior and
causing unnecessary kernel respecialization via ArraySpec type params.

Author: maleadt

src/language/types.jl

@@ -79,7 +79,7 @@ Returns 0 if value is 0 or not divisible by any power of 2.
 function compute_divisibility(value::Integer, max_divisor::Int=16)
     value == 0 && return 0
     divisor = 1
-    while divisor <= max_divisor && value % (divisor * 2) == 0
+    while divisor < max_divisor && value % (divisor * 2) == 0
         divisor *= 2
     end
     return divisor >= 2 ? divisor : 0  # Only return if at least divisible by 2