Initial pool size should not be hard coded

[harrism]

Now that rmm::pool_memory_resource does not by default use half of available memory for the initial pool size, cuDF has hard-coded 50% of free memory as the initial size to match previous behavior for benchmarks and tests. This should be configured globally somewhere.

         I'm thinking about whether we should parse `percent_of_free_device_memory` from some env variable, instead of hard coding like this in many places. By doing so we can control the number we want at runtime.

Originally posted by @ttnghia in #14741 (comment)