parallel download support

[lmeyerov]

In benchmarks, we're finding significant speedups by using parallel downloads for MB+ individual blobs:

for a 200MB file on an az server <> az store, warm:

• fs.download: 4.9s
• file.write(await (await bc.download_blob()).readall())): 4.3s
• file.write(await (await bc.download_blob(max_concurrency=16)).readall())): 1.3s
• trickier: I think it's even bigger in practice b/c above was a warm read. cold read baselines were more 10-20s.

This is on a node with 8 cores and 2 NICs, and I think < 1 Gbps

Ideally:

• smart default: concurrency defaults to # cores x2 or x4, something like that
• unclear: handling of recursive case
• allow manual overrides
• works with dask, e.g., dd.read_parquet(...)

[hayesgb]

@lmeyerov -- Really appreciate this benchmarking! I've put together a crack at this. It's currently implemented in the concurrent_io branch for:

fs.get_file()
fs.put_file()
fs.open()
AzureBlobFile.write()

When I timed this on my local machine, put_file gave me over 50% reduction in write times, so I'd appreciate any feedback you have.