Update the readme and fix bugs in custom-dataset example

[lancerts]

1. Update the Caltech 101 dataset URL in readme.
2. random_shuffle: is not a member of 'std' since it c++17, updated to shuffle.
3. Functional(torch::log_softmax, 1, torch::nullopt) no longer matches the constructor, updated to lambda expression.
4. auto tlabel = torch::from_blob(&data[index].second, {1}, torch::kLong); sometimes leads to non-sensical numerical value.
   A few examples from std::cout:

data[index].second: 42 
tlabel:  42 [ CPULongType{1} ] 
data[index].second: 10 
tlabel:  10 [ CPULongType{1} ] 
data[index].second: 98 
tlabel:  4.6698e+16

Updated to direct the creation of a tensor instead of creation from a pointer.

Tested the program via libtorch (2.1.2) with cuda 12.1 using both the release and debug build.

[netlify]

✅ Deploy Preview for pytorch-examples-preview canceled.

Name	Link
🔨 Latest commit	fb0ecc7
🔍 Latest deploy log	https://app.netlify.com/sites/pytorch-examples-preview/deploys/65a1b7b2ee76ec00087bc152