[BACKEND] bump to most recent LLVM commit 2fe947b477

[makslevental]

I have "bisected" the current CUDA regression (mentioned here) on A100:

git bisect start
# bad: [f5a35a31bfe6cbc16bec0c130f2bb3632dbf1fbf] [LV] Add test cases with incorrect IV live-outs.
git bisect bad f5a35a31bfe6cbc16bec0c130f2bb3632dbf1fbf
# good: [86b69c31642e98f8357df62c09d118ad1da4e16a] [SPIR-V] Fix SPIR-V extension SPV_INTEL_function_pointers: 
...
# first bad commit: [b8703369daf777706196ff914c0376c27adde3cf] [SLP] Match poison as instruction with the same opcode

The failure presents as :

Traceback (most recent call last):
  File "/home/mleventa/dev_projects/repro/repro.py", line 135, in <module>
    torch.cuda.synchronize()
  File "/home/mleventa/dev_projects/.venv/lib/python3.10/site-packages/torch/cuda/__init__.py", line 987, in synchronize
    return torch._C._cuda_synchronize()
RuntimeError: CUDA error: an illegal memory access was encountered

The failing commit is b870336 and the last working commit is 2fe947b4. It's not clear to me the relationship but will investigate.

This PR to llvm-head is to verify that my repro on A100 matches the internal regression/fail.

[makslevental]

FYI this is the reproducer used https://gist.github.com/jataylo/5ebe20d3fac7a84cb24e69c0b7b449f5 (credit to @jataylo)

[makslevental]

Sanity check: putting tuningOptions.SLPVectorization = false; "unblocks" the regression (I.e., the reproducer passes on LLVM HEAD).