Add "blobs" to the table format

[westonpace]

We are currently writing large values such as images and videos into the the data files themselves. Since the format is designed to handle large values this works reasonably well. However, for very large values (4/8 MB and above) this causes issues with compaction (both RAM required to run compaction and the total amount of I/O that must be performed).

We can work around this by introducing blob files. These files contain blob data. The data files contain a pointer to the blob file. When fragments are compacted the blob files are not necessarily compacted.

Blob files may be compacted if they are small to hit a target size. Once a blob hits a target size it is no longer compacted (except possibly to materialize deletions but this is tricky). When blobs are compacted then affected fragments need to have their blob pointers remapped.

Blobs also need to be supported on the cleanup path. Once a blob file is no longer referenced by any fragments that blob file should be removed.

Tasks:

• Initial support for writing blob files #2267
• Support for blob compaction (and the needed remapping) #2268
• Store blob paths with dictionary encoding
• Support for blob file cleanup #2270
• Support for materialized deletion of blob files (if needed)
• Document blobs in the table format docs
• Add tutorial / example for blobs in python docs