[FEA] Evaluate exposing CAGRA graph size as a `long`

Background

#1086 added getters for CAGRA index graph sizes/degrees/data-pointers via the C-API, for use from Java (etc.).

One sees that the graph-size is exposed as a uint32_t: Crucially, unsigned.

The problem

The challenge here is that Java has no support for unsigned types. When the graph-size is fetched via JNI (Panama), it is exposed as a signed int type. If used with a graph whose size exceeds 2**31, the Java code will likely run with an incorrect negative value. (First reported here.)

The workaround

The Java side currently addresses this via an up-cast to long.

Preferred solution:

It would be good to evaluate whether the graph-size should really be exposed as a long in the C API. It would save having to cast on the receiving end.
This would also be in line with how the number of rows in the dataset is exposed: as a long.

Review the C-API graph-size getters introduced by #1086 and the Java-side up-cast in PR #1216. Compare the graph-size type with the dataset row-count API, then determine whether changing the exposed type is safe for the C API and Java consumers. Done means values above 2**31 remain correct without relying on an unsigned-to-signed conversion.