aot_eager amp training failure: conversion is supported for Half type only

Discovered in a recent pytorch nightly pin update, https://github.com/pytorch/torchdynamo/pull/1351

Repro:
```
benchmarks/torchbench.py --accuracy --training --amp --backend=aot_eager --only pytorch_struct
```
Error:
```
TorchDynamo optimized model failed to run because of following error
conversion is supported for Half type only
FAIL
```