Removing deprecated `cpp_custom_type_hack`

[peterbell10]

Context

at::cpp_custom_type_hack is a deprecated utility to shepherd arbitrary C++ classes through Python by embedding an owning pointer inside a byte Tensor's storage. It's usage is already heavily discouraged:

pytorch/aten/src/ATen/cpp_custom_type_hack.h

Lines 18 to 20 in 2291981

	// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
	// YOU ARE IN THE WRONG PLACE! TURN BACK NOW!

torch/custom_class.h is now the preferred way to expose C++ classes to Python, however the current uses may be serialized into existing TorchScript models so these operators must be changed with forward and backward compatibility in mind.

Relevant operators

• aten::quantized_gru.input_legacy

• aten::quantized_gru.data_legacy

• aten::quantized_lstm.input_legacy

• aten::quantized_lstm.data_legacy

• aten::fbgemm_linear_int8_weight_fp32_activation

• aten::fbgemm_pack_quantized_matrix

• aten::fbgemm_pack_gemm_matrix_fp16

• aten::fbgemm_linear_fp16_weight_fp32_activation

• quantized::linear_prepack_legacy

• quantized::linear_prepack_fp16_legacy

• _quantized::linear_prepack_fp16_legacy

• profiler::_record_function_enter (see Refactor record_function_ops.cpp to not use cpp_custom_type_hack #35026)

• profiler::_record_function_exit (see Refactor record_function_ops.cpp to not use cpp_custom_type_hack #35026)

cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @Xia-Weiwen @leslie-fang-intel @robieta @chaekit @aaronenyeshi @nbcsm @guotuofeng @guyang3532 @gaoteng-git @tiffzhaofb @dzhulgakov @davidberard98 @ilia-cher @gdankel @bitfort @ngimel @orionr @albanD @malfet