Use bst_ulong instead of size_t for CSR conversion

[yzhliu]

Some platform, e.g., WebAssembly, has strict typing checking. The mismatch will cause crash.

And they should be size_t to be consistent with https://github.com/dmlc/xgboost/blob/master/src/c_api/c_api.cc#L392 and https://github.com/dmlc/xgboost/blob/master/src/data/proxy_dmatrix.cc#L19

[trivialfis]

Thank you for the fix! But can we keep the use of bst_ulong? It's defined as uint64_t, which is more strict than size_t. There's c_bst_ulong in the _typing.py module.

[yzhliu]

@trivialfis I'm ok with changing python side to c_bst_ulong. But do we really want to do that? cuz all the downstreams are size_t, examples like https://github.com/dmlc/xgboost/blob/master/src/data/adapter.h#L352 and ArrayInterface.Shape() used by it. IMO using size_t would be more general that works for old 32bit machines and future 128bit machine?

not strong opinion though.

[trivialfis]

I think we will have improved consistency (which is currently lacking for XGBoost) with other C API functions if we use uint64_t. Using uint64_t here to record the shape of the dataset doesn't have practical implications for different memory address sizes. We can optionally define bst_ulong as uint128_t if the future Skynet decides that having a single memory address space to index exabytes of memory is necessary. ;-)

A side note, c++ ptrdiff_t is usually defined as int64_t on 64-bit machines, as a result, that's the practical limit of memory indexing size.

[trivialfis]

Thank you for the fix!