[FEA] Improve cudf::io::datasource::create()

At a prior company that made heavy use of cudf, we ran into serious performance problems with Kvikio/GDS when small reads were being furnished by a mmap'd data source because they did not meet the configured device read threshold (i.e. for reading compressed metadata which would frequently be less than 128KB).

This was because the underlying parquet files lived on a network file system--WekaFS in our case--and mmap'd I/O against WekaFS resulted in pathological performance problems. Especially when it was lots of little <128KB reads that would be paged in 4K at a time over the network in a really inefficient way.

Additionally, we found ourselves needing to have a data source that did O_DIRECT host reads for everything, such that the Linux page cache could be avoided entirely. This was useful when doing back-to-back benchmarks of different features, where the presence or absence of data in the page cache would have huge impacts on runtimes. i.e. it was impossible to see if a PR purporting a 5% perf boost was actually delivering such a boost when back-to-back runs varied by 50% simply due to cold vs hot page cache data.

I'm working on reimplementing the various improvements we introduced in our internal codebase in cudf, such that they can be used by others. The approach that seems most viable, assuming you want to keep all datasource implementation innards behind `datasource.cpp`, is to allow `cudf::io::datasource::create()` to take additional `datasource_kind` and `datasource_params` parameters; see code sample below.

I'm actively hacking on a branch locally, I'll throw up a draft PR soon.

Edit: PR is active here: https://github.com/rapidsai/cudf/pull/17115. Removing the temp header diff pasted initially.