Comment out eliminate-empty-tensors (#1866)

**Context:** The compile time increased too much after the migration to
the one-shot bufferization pipeline. The bufferization
[documentation](https://mlir.llvm.org/docs/Bufferization/) recommends
running -eliminate-empty-tensors before running the one-shot
bufferization. However, eliminating empty-tensors takes half of the
already expensive compilation time for programs containing large
functions.

Without a clear benefit of running eliminate-empty-tensors and a clear
cost in compilation time for programs of interest, commenting out this
pass will drop compile times significantly in the programs of interest.

**Description of the Change:** Comment out the -eliminate-empty-tensor
pass from the frontend to avoid running it.

**Benefits:** Decreased compilation time.

**Possible Drawbacks:** Possible (unmeasured) increase in runtime memory
allocations.

**Related GitHub Issues:**

Part 1 of [sc-93707]

Co-authored-by: Joseph Lee <[email protected]>

Author: erick-xanadu

frontend/catalyst/pipelines.py

@@ -274,7 +274,10 @@ def get_bufferization_stage(options: CompileOptions) -> List[str]:
         "convert-tensor-to-linalg",  # tensor.pad
         "convert-elementwise-to-linalg",  # Must be run before --one-shot-bufferize
         "gradient-preprocess",
-        "eliminate-empty-tensors",
+        # "eliminate-empty-tensors",
+        # Keep eliminate-empty-tensors commented out until benchmarks use more structure
+        # and produce functions of reasonable size. Otherwise, eliminate-empty-tensors
+        # will consume a significant amount of compile time along with one-shot-bufferize.
         ####################
         "one-shot-bufferize{" + bufferization_options + "}",
         ####################