[NO CP][ROCm][inductor][release/2.9] Fix tanh for FP64

[naromero77amd]

fast_tanhf does not support float64 and gives a compiler error. fallback to tanh for float64.

Will fix SWDEV-557937 once it is backported to release_2.8.

Related to:
pytorch#162052

[rocm-repo-management-api]

Jenkins build for 5a6d97ed21b12564251f03d71a008a390fc8b402 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

[jataylo]

Approved but lets wait for your test suite results.

[naromero77amd]

Ran tests as follows:

PYTORCH_TEST_WITH_ROCM=1 TEST_CONFIG=default python test/run_test.py --continue-through-error --verbose -i inductor/test_torchinductor.py  inductor/test_max_autotune.py \
    inductor/test_op_dtype_prop.py inductor/test_torchinductor_opinfo.py >& ../$logfile

I used chatgpt to analyze the logs. The following tests are now passing:

test/inductor/test_op_dtype_prop.py::TestCaseCUDA::test_dtype_aw...are_codegen_op_name_tanh_load_upcast_to_fp32_False_bfloat16_cuda
test/inductor/test_op_dtype_prop.py::TestCaseCUDA::test_dtype_aw...tanh_load_upcast_to_fp32_False_float16_cuda
test/inductor/test_op_dtype_prop.py::TestCaseCUDA::test_op_dtype_propagation_tanh_cuda_float64

test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_gelu_cuda_float64
test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_mish_cuda_float64
test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_tanhshrink_cuda_float64
test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_tanh_cuda_float64

there are no new regressions.