FIX Reset the FP32 matmul precision in tests (#2411)
Fixes currently failing hotswap+compile tests that fail because outputs
are not close enough before vs after compilation.
In test_gpu_examples.py, some tests run torchao, which sets the float32
matmul precision to "high". This in turn results in some models
producing different outputs when compiled (but only for some seeds).
Therefore, we need to ensure that the precision is reset to "highest",
which is the default.
