Qualcomm AI Engine Direct - Add MHA2SHA pass (pytorch#15438)

### Background
We observed that quantizing and compiling the original sha model
requires a significant amount of time. Switching to the mha model speeds
up this process. Therefore, we investigated whether converting the mha
model after quantization is feasible. However, we cannot perform this
conversion during the to_edge transformation, as splitting the
convolution weights to sha would require modifying the state_dict, which
is not permitted at that stage. Therefore, we decided to apply this pass
during qnn_preprocess.

### Summary:
- Integrated mha into sha pass and implemented it in qnn_preprocess
- Refactored mha in static llama
- Included spin quant r3 support and masked softmax for MHA model in
static llama
- Combined the n_heads key-value cache into a single cache for each
layer to decrease the number of inputs and outputs, which enhances
performance.
- Deprecated ShiftPointer kv updater mode
- Since each layer now has its own kv cache, the v cache no longer
benefits from ShiftPointer, which previously avoided copying the new v
cache to the input v cache. To prevent user confusion, ShiftPointer mode
has been deprecated
- Applied the correct input template for smollm2 135m
- Correct the quantization annotation for reshape
- Remove outdated code from CanonicalizeConv


### Results
Follow [README
setting](https://github.com/pytorch/executorch/blob/main/examples/qualcomm/oss_scripts/llama/README.md),
test on SM8750 with QNN 2.37. Compared the new pass `convert_mha_to_sha`
with original sha structure
 
<img width="1731" height="734" alt="image"
src="https://github.com/user-attachments/assets/2b1c2b66-77c0-4662-a035-900ad9091d67"
/>

Author: shewu-quic

backends/qualcomm/_passes/__init__.py

@@ -11,6 +11,7 @@
 from .canonicalize_conv import CanonicalizeConv
 from .convert_bmm_to_matmul import ConvertBmmToMatmul
 from .convert_linear_to_conv2d import ConvertLinearToConv2d
+from .convert_mha_to_sha import ConvertMhaToSha
 from .convert_square_to_pow import ConvertSquareToPow
 from .decompose_any import DecomposeAny
 from .decompose_binary_alpha import DecomposeBinaryAlpha
@@ -56,6 +57,7 @@
     CanonicalizeConv,
     ConvertBmmToMatmul,
     ConvertLinearToConv2d,
+    ConvertMhaToSha,
     ConvertSquareToPow,
     DecomposeAny,
     DecomposeBinaryAlpha,

backends/qualcomm/_passes/canonicalize_conv.py

@@ -9,7 +9,6 @@
 import torch
 
 from executorch.backends.qualcomm.builders.utils import get_parameter, set_parameter
-from executorch.backends.qualcomm.utils.constants import QCOM_REQUANTIZE
 from executorch.exir.pass_base import ExportPass, PassResult
 from torch._guards import detect_fake_mode
 
@@ -197,14 +196,6 @@ def call(self, graph_module: torch.fx.GraphModule):
                             )
                             squeeze_node.meta = copy_meta(node.meta)
 
-                            if QCOM_REQUANTIZE in input_node.meta:
-                                input_node.meta.pop(QCOM_REQUANTIZE)
-                            if QCOM_REQUANTIZE in node.meta:
-                                squeeze_node.meta[QCOM_REQUANTIZE] = node.meta[
-                                    QCOM_REQUANTIZE
-                                ]
-                                conv2d_node.meta.pop(QCOM_REQUANTIZE, None)
-
                 for user in node.users.copy():
                     user.replace_input_with(node, squeeze_node)
 

backends/qualcomm/_passes/convert_bmm_to_matmul.py

@@ -47,7 +47,13 @@ def call(self, graph_module: torch.fx.GraphModule):
         graph = graph_module.graph
         partitions = get_source_partitions(
             graph,
-            [operator.matmul, torch.matmul, torch.bmm, torch.ops.aten.matmul.default],
+            [
+                "matmul",
+                operator.matmul,
+                torch.matmul,
+                torch.bmm,
+                torch.ops.aten.matmul.default,
+            ],
         )
         for _, src_partitions in partitions.items():
             for src_partition in src_partitions: