Update on "[ET-VK] Implement SDPA with fused ops"

## Context

As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps:

1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask
2. Compute softmax normalization of computed attention weights
3. Compute final output by multiplying attention weights with V cache

This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped.

## Impact

Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s

Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/)

[ghstack-poisoned]

Author: ssjia

.github/workflows/pull.yml

@@ -973,7 +973,10 @@ jobs:
 
         # "Classic" Operator tests
         PYTHON_EXECUTABLE=python bash backends/vulkan/test/scripts/test_op.sh --build
-        ./cmake-out/backends/vulkan/test/op_tests/vulkan_sdpa_test
+        # TODO(ssjia): figure out how to run custom op tests in CI. Currently, they are
+        # failing due to to the libstdc++.so.6 installed with conda not supporting
+        # GLIBCXX_3.4.30. These tests are still run in Meta internal CI.
+        # ./cmake-out/backends/vulkan/test/op_tests/vulkan_sdpa_test
 
         # Run e2e testing for selected operators. More operators will be tested via this
         # route in the future.