Port Hornet's binning block memory allocator to cuGraph

[seunghwak]

This PR ports Hornet’s dynamic graph memory allocator into cuGraph under cugraph/cpp/include/cugraph/dynamic/memory_manager. This is a first task to support vertex/edge insertions/deletions in cuGraph.

The allocator has three layers:

bit_tree_t — Host-side Vec-Tree bitmap (from Hornet). Tracks which fixed-size blocks are free inside one block array. insert() returns a block index; remove(block_index) frees it back.

block_array_t — One GPU buffer (rmm::device_uvector for arithmetic types or cugraph::dataframe_buffer_t for cuda::std::tuple of arithmetic types) plus a bit_tree_t. The buffer holds the actual data; the bit tree hands out fixed-size blocks within it. insert()/remove() allocate and free blocks.

block_array_manager_t — Maintain separate block array pools based on the requested block size. Block size in different pools are power of 2 numbers (2^0, 2^1, 2^2, ... ). Given a requested block size, it finds a non-full block array in the matching pool or creates a new one. insert() returns block_access_data_t (block_array_key, block_index, num_elements_per_block); remove() uses that handle to free the block.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[seunghwak]

/ok to test c1d7e35

[seunghwak]

/ok to test 46aa306

[seunghwak]

/ok to test 8e0e92f

[ChuckHastings]

/merge