Introduce I64toI32 export pass (#3727)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/3727

## Context

A number of `nn.Module`s targeting the Vulkan delegate use `i64` dtype for operators, inputs, and outputs. This is because `i64` is the default for many `torch` functions. Since [`i64` dtype is a Vulkan extension](https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VK_KHR_shader_atomic_int64.html), it is not supported on all Vulkan devices. Hence, we introduce an export pass that converts the majority of the graph to `i32` dtype:
1. For each `i64` dtype input, append a node to convert the output tensor to `i32` dtype.
2. For each operator yielding `i64` dtype, replace it with `i32` dtype.
3. For each `i64` dtype output, prepend a node to convert the input tensor to `i64` dtype.

## Example
Take this simple model compiled to one Clamp operation with input `x = (torch.randint(low=-100, high=100, size=(5, 5)),)`. By default, `torch.randint` uses `i64` dtype.
```
class ClampModule(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        x = torch.clamp(x, min=-3)
        return x
```

If it's compiled with such `i64` input, the export pass rewrites the graph to be the equivalent of the following.
```
class GoalModule(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        x = x.to(torch.int32)
        x = torch.clamp(x, min=-3)
        x = x.to(torch.int64)
        return x
```

ghstack-source-id: 229431025
exported-using-ghexport

Reviewed By: SS-JIA

Differential Revision: D57649650

fbshipit-source-id: ba36f1a563c2d97021b8f7d1ff1929d29223c43d