Add support for single-step write to S3, instead of multi-part

[ihnorton]

This PR adds support for writing to an S3-compatible store with the PutObject method, rather the multi-part upload API. Primarily for compatibility with GCS, where the S3 compatibility mode
does not support multi-part uploads. Path is controlled by the new config parameter s3.use_multipart_upload, with default true.

[ihnorton]

This is ready for review.

I've tested it with arrays up to 1GB on GCS, and it seems to work fine. There is a configuration override required currently -- in order to use a buffer that is larger than the ~40MB available with config defaults, I set these options:

# set the maximum upload buffer to 50GB
config["vfs.s3.multipart_part_size"] = int(5e10)
config["vfs.s3.max_parallel_ops"] = 1

This could be an additional, separate config option, or we could consider:

• leave it as-is while the feature shakes out, rather than committing to a new config option now
• renaming multipart_part_size to (something like) max_upload_size and document it as either per-chunk-max or direct-write-max depending on whether use_multipart_upload is enabled.

Noting the limits of various services here:

• AWS has a PUT maximum of 5GB
• As far as I can tell, GCS objects are only subject to a maximum of 5TB (no individual limit per-upload)

[tdenniston]

LGTM -- it might be worth adding a GCS section to https://docs.tiledb.io/en/stable/tutorials/working-with-s3.html to document this. I think sticking with the existing config param is fine for now.

[ihnorton]

will rebase on #1238 if that passes CI.