Add mmap based storage of vectors, while keeping buffered storage as default

[rchitale7]

Description

This issue enables the vector and doc id binary downloaded from Remote Store to be saved to disk, and read using mmap. The caller of run_tasks can download the vectors to disk if they set storage_mode = "disk", but the default storage mode is still memory. I refactored the code so that we can support other storage modes if necessary. Key to the refactoring is the BinarySource object that wraps the storage mechanism; FileSource wraps the file object used for numpy.mmap, while BufferSource wraps the buffer object used for numpy.frombuffer.

I verified manually that both storage modes work, and create the expected faiss graph.

Issues Resolved

Partially resolves for #49. I need to still update the USER_GUIDE.md to specify that memory is the default approach, but with CUDA versions >= 12, we see memory spike. So to avoid memory spike, user can use disk (at the expense of slower index builds)

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.
For more information on following Developer Certificate of Origin and signing off your commits, please check here.

[rchitale7]

PR is pending resolution of the discussion here: facebookresearch/faiss#4274

[rchitale7]

Based on discussion in facebookresearch/faiss#4274, we determined that the reason the memory spikes when storing vectors in CPU memory is because of the CUDA version. This issue is observed with CUDA version >= 12. I've updated this PR to make memory based storage the default, while still giving the caller the option to use disk/mmap based storage as another option. I'll update the USER_GUIDE.md to specify when to use disk v.s. memory storage after this PR is merged.

[navneet1v]

@rchitale7 fix the commit message and also the PR description to correctly point out what the PR is doing

[rchitale7]

@rchitale7 fix the commit message and also the PR description to correctly point out what the PR is doing

missed this, fixed it now.

[navneet1v]

@rchitale7 Based on the response here: facebookresearch/faiss#4274 (comment) I think we might not even need this PR. So, I will wait for the bug to be fixed.