Create specific TTNN op for empty tensor on host

[vwellsTT]

Problem

As a quality-of-life improvement for cpu hoisting, I want a way to lower tensor.empty to on host. Currently, TTNN::EmptyOp does not work because the tt_metal api for it requires a device and memory_config which don't have meaningful values for on-host tensors. Existing tests workaround the issue by using say ttnn.zeros instead, but this is a bit clunky--I propose an op to directly invoke a tensor ctor instead.

Changes

Create new ConstructTensor op which maps to Tensor ctor invocation at runtime. Lower tensor.empty ops in system_memory into this instead of ttnn.empty_op + add support in flatbuffer and runtime

Example

input

  func.func @add(%arg0: tensor<32x32xf32>, %arg1: tensor<32x32xf32>) -> tensor<32x32xf32> {
    %0 = tensor.empty() : tensor<32x32xf32>
    %1 = "ttir.add"(%arg0, %arg1, %0) <{operandSegmentSizes = array<i32: 2, 1>}> {should_hoist} : (tensor<32x32xf32>, tensor<32x32xf32>, tensor<32x32xf32>) -> tensor<32x32xf32>
    return %1 : tensor<32x32xf32>
  }

output

func.func @add(%arg0: tensor<32x32xf32, #ttnn_layout>, %arg1: tensor<32x32xf32, #ttnn_layout>) -> tensor<32x32xf32, #ttnn_layout> {
        // input data movement
        %5 = "ttnn.construct_tensor"() <{dtype = #tt.supportedDataTypes<f32>, layout = #ttnn.layout<row_major>, shape = #ttnn.shape<32x32>}> : () -> tensor<32x32xf32, #ttnn_layout2>
        %6 = call @hoisted_ttir_add_32x32_32x32_32x32_func_decl(%2, %4, %5) : (tensor<32x32xf32, #ttnn_layout2>, tensor<32x32xf32, #ttnn_layout2>, tensor<32x32xf32, #ttnn_layout2>) -> tensor<32x32xf32, #ttnn_layout2>
        // output data motion
        return %8 : tensor<32x32xf32, #ttnn_layout>
      }