triton: type conversion before tl.dot fails compilation

Context: we are trying to upgrade trition pin in pytorch and start encountering test failure for python test/inductor/test_pattern_matcher.py -k test_mixed_mm

Doing a type conversion for an input tensor to tl.dot starts to fail. It works in triton commit: e6216047b8b0aef1fe8da6ca8667a3ad0a016411 , but fail on a recent commit 0410652666ddcad54aa4c403276b478a36379b90 .

Here is a standalone repro w.py: https://gist.github.com/shunting314/3a3b8ce1ccee7b51b8ee0d9a2d24dd3d Running python w.py will report the following error in the new commit:

loc("/tmp/torchinductor_shunting/bm/cbm7qsh5esh6xdkdddmv7l2ilel4kdbfwgy2luolzmme62njagrb.py":64:17): error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast
Failed to emit LLVM IR
Translate to LLVM IR failedLLVM ERROR: Failed to translate TritonGPU to LLVM IR.
Aborted (core dumped)

Here is a same repro but with inductor dependencies removed: https://gist.github.com/shunting314/4eb4a6a7d8cbfc6396726518196e26d1