Update on "[ET-VK] Conv2d quantize/dequantize ops for conv2d activations"

## Context

As title. Add shaders to quantize a floating point conv2d input tensor to packed int8 memory layout and dequantize a int8 conv2d output tensor back to floating point representation.

Hooking it up to the export logic will be handled in a follow up diff.

Differential Revision: [D82542335](https://our.internmc.facebook.com/intern/diff/D82542335/)

[ghstack-poisoned]

Author: ssjia

backends/vulkan/test/custom_ops/CMakeLists.txt

@@ -95,4 +95,5 @@ if(TARGET vulkan_backend)
   add_operator_prototype(q8csw_conv2d)
   add_operator_prototype(q4gsw_linear)
   add_operator_prototype(choose_qparams_per_row)
+  add_operator_prototype(qdq8ta_conv2d_activations)
 endif()

backends/vulkan/test/custom_ops/qdq8ta_conv2d_activations.cpp

@@ -9,7 +9,6 @@
 #include <algorithm>
 #include <cmath>
 #include <iostream>
-#include <random>
 #include <vector>
 #include "utils.h"
 
@@ -58,15 +57,13 @@ TestCase create_test_case_from_config(
       input_dtype,
       storage_type,
       utils::kChannelsPacked, // Use channels packed for conv2d tensors
-      DataGenType::RANDINT8);
+      DataGenType::RANDOM);
 
-  // Randomized quantization parameters following q8csw_conv2d.cpp pattern
-  // Generate random scale value
-  float scale_val = 1.0;
+  float scale_val = 0.007112;
   ValueSpec scale(scale_val);
 
   // Generate random zero point within quantization range
-  int32_t zero_point_val = 0;
+  int32_t zero_point_val = -2;
   ValueSpec zero_point(zero_point_val);
 
   // Output tensor (float) - same shape as input [N, C, H, W]