[06-fused-attention] Temporarily modify the source code to get back performance (#4399)

This PR relands recent commits to tutorial 06-fused-attention, and
performs the minimal changes in the source code to get back the original
performance.

@etiotto identified the cause of the FP16 performance regression is due
to different implementation of transpose in the source code:
#4283 (comment).
This PR revert back to the old implementation. The last commit of this
PR should be reverted when the new implementation of transpose can be
handled efficiently.

Fixes #4283

Author: whitneywhtsang

Makefile

@@ -30,8 +30,8 @@ test-cpp:
 
 .PHONY: test-unit
 test-unit: all
-	cd python/test/unit && $(PYTEST) -s -n 8 --ignore=cuda/test_flashattention.py \
-		--ignore=language/test_line_info.py --ignore=language/test_subprocess.py --ignore=test_debug.py
+	cd python/test/unit && $(PYTEST) -s -n 8 --ignore=language/test_line_info.py \
+		--ignore=language/test_subprocess.py --ignore=test_debug.py
 	$(PYTEST) -s -n 8 python/test/unit/language/test_subprocess.py
 	$(PYTEST) -s -n 8 python/test/unit/test_debug.py --forked
 	$(PYTEST) -s -n 8 python/triton_kernels/tests/