[FEA] Java: Add ai.rapids.cudf.HashJoin.getDeviceMemorySize introspection

**Is your feature request related to a problem? Please describe.**

Allow applications to understand how much GPU device memory they are retaining though cudf objects like `HashJoin`.
**Describe the solution you'd like**

- Add `getDeviceMemorySize` or similar method to `ai.rapids.cudf.HashJoin` similar to existing `ai.rapids.cudf.Table.getDeviceMemorySize`
- when the caller holds the source `Table` and the `HashJoin` built from that table, it should have a way to estimate total retained memory usage without double-counting shared columns, if any

**Describe alternatives you've considered**

Not knowing retaining memory. However this make it hard from utilizing available memory smartly.

**Additional context**

None

Start with the Java API entry points ai.rapids.cudf.HashJoin and ai.rapids.cudf.Table.getDeviceMemorySize, then trace how HashJoin retains device memory. Define how the new introspection method estimates total retained memory when source Tables share columns, without double-counting them. Done means callers can query HashJoin-related GPU memory consistently with Table introspection.

Written by the indexing model from the issue text.