[Cherry-pick] Enable cross entropy loss for xla autocast with FP32 precision (#8094) #10114

Job	Run time
get-torch-commit	2s
Build XLA CUDA plugin / build	5m 29s
Build PyTorch/XLA / build	33m 16s
Build PyTorch with CUDA / build	24m 5s
TPU tests	0s
GPU tests / test (benchmark_tests)	13m 46s
GPU tests / test (python_tests, xla_op1)	32m 1s
GPU tests / test (python_tests, xla_op2)	8m 32s
Build docs / build-docs	2m 52s
GPU tests / test (python_tests, xla_op3)	44m 48s
GPU tests requiring torch CUDA / test (python_tests, linux.8xlarge.nvidia.gpu)	10m 55s
GPU tests requiring torch CUDA / test (triton_tests, linux.g5.4xlarge.nvidia.gpu)	8m 26s
GPU tests / test (python_tests, torch_mp_op)	30m 7s
GPU tests / test (cpp_tests, cpp_tests1)	11m 0s
GPU tests / test (cpp_tests, cpp_tests2)	18m 12s
CPU tests / test (benchmark_tests)	7m 56s
CPU tests / test (python_tests, xla_op1)	15m 25s
CPU tests / test (python_tests, xla_op2)	4m 39s
CPU tests / test (python_tests, xla_op3)	7m 0s
CPU tests / test (python_tests, torch_mp_op)	10m 24s
CPU tests / test (cpp_tests, cpp_tests1)	6m 18s
CPU tests / test (cpp_tests, cpp_tests2)	4m 55s
	5h 0m 8s

Provide feedback