[llm integrate] wip faster codegen and various improvements

[sklam]

• making pytest ./llm_midend.py all passes
• strip trailing whitespaces

[kc611]

The following patch fixes the current test failures:

diff --git a/sealir-tutorials/llm_midend.py b/sealir-tutorials/llm_midend.py
index ef9c8a5..d45b3c3 100644
--- a/sealir-tutorials/llm_midend.py
+++ b/sealir-tutorials/llm_midend.py
@@ -3003,7 +3003,7 @@ class MlirBackend(_ch06_MlirBackend):
                     ir.AffineMap.get_identity(nd),
                     ir.AffineMap.get_identity(nd)
                 ],
-                iterator_types=ir.ArrayAttr.get([ir.Attribute.parse("#linalg.iterator_type<parallel>")]*2)
+                iterator_types=ir.ArrayAttr.get([ir.Attribute.parse("#linalg.iterator_type<parallel>")]*nd)
             )
 
             body = generic_op.regions[0].blocks.append(
@@ -3039,7 +3039,7 @@ class MlirBackend(_ch06_MlirBackend):
                 ir.AffineMap.get_identity(nd),
                 ir.AffineMap.get_identity(nd)
             ],
-            iterator_types=ir.ArrayAttr.get([ir.Attribute.parse("#linalg.iterator_type<parallel>")])
+            iterator_types=ir.ArrayAttr.get([ir.Attribute.parse("#linalg.iterator_type<parallel>")]*nd)
         )
 
         body = generic_op.regions[0].blocks.append(

[sklam]

By 50c82ff,
I'm starting to test tensor dialect in an attempt to get optimal code.