[LANG] Added support for constexpr (#361)

python/tutorials/02-fused-softmax.py

@@ -65,11 +65,11 @@ import triton.language as tl
 
 @triton.jit
 def softmax_kernel(
-    output_ptr, input_ptr, input_row_stride, output_row_stride, n_cols, **meta
+    output_ptr, input_ptr, input_row_stride, output_row_stride, n_cols, 
+    BLOCK_SIZE: tl.constexpr
 ):
     # The rows of the softmax are independent, so we parallelize across those
     row_idx = tl.program_id(0)
-    BLOCK_SIZE = meta['BLOCK_SIZE']
     # The stride represents how much we need to increase the pointer to advance 1 row
     row_start_ptr = input_ptr + row_idx * input_row_stride
     # The block size is the next power of two greater than n_cols, so we can fit each