Use PG cursors + FETCH to partition tables on commit.

Instead of doing a pre-flight query to first figure out partition boundaries,
then chunking the table by filtering on those boundaries (poor performance on
tables without PKs and foreign tables), use a PG server-side cursor and run
FETCHes to add data to CStore files.

Two changes to behaviour: not making objects in parallel and moving the object
hashing to be after we've grabbed the data from the cursor into a temporary
object (after which we actually rename it).

Other simplifications to object creation:

When renaming the object, do a pre-flight check to make sure it doesn't
already exist instead of setting up savepoints. If we do get a race, it's fine
because we'll just reupsert existing meta(data) with the same values.

Also simplify create_base_fragment (delete now-unneeded chunking logic)

Benchmarking

Setup: I had a CSV file with some data after running sgr cloud download:

$ zcat query-e2abcf218bbc964ff9999b08c06c97447018c395-20211217-134600.csv.gz | head
timestamp,query_id,remote_address,application,user_id,username,matomo_id,access_token,dbname,original_query,success,used_images
2020-08-27 21:30:24,--XUMXQB-GrJfm60NHBY,82.4.198.115,pgcli,7575f36b-5f6e-4f86-b05e-b1cc91c73fa1,mildbyte,,false,,"SELECT * FROM ""splitgraph/election-geodata:latest"".nation g
WHERE g.state = '36'
        AND g.county = '061' AND g.precinct = '3606183003'
        AND ST_IsValid(geom)",t,"[""splitgraph/election-geodata:latest""]"
$ zcat query-e2abcf218bbc964ff9999b08c06c97447018c395-20211217-134600.csv.gz | wc -l
12232602


$ sgr init benchmark/commit-speed
$ zcat query-e2abcf218bbc964ff9999b08c06c97447018c395-20211217-134600.csv.gz | sgr csv import benchmark/commit-speed query-results
Warning: primary key is not specified, using the whole row as primary key.This is probably not something that you want.

Without a primary key

time sgr commit --snap --overwrite benchmark/commit-speed

Old sgr: started a tqdm progressbar with a high-variance ETA (because objects are committed in parallel, so we get 8 objects bumping the pbar at the same time). Terminated after 25 minutes with tqdm estimating a total time of 1h30m.

New sgr: 3 trials, 5m36s, 5m3s, 5m29s (avg 5m23s)

With a primary key

sgr sql -s benchmark/commit-speed 'ALTER TABLE "query-results" ADD PRIMARY KEY (timestamp, query_id)'

Old sgr: 5m32s, 6m14s, 5m49s (avg 5m52s)
New sgr: 3m23s, 3m39s, 3m15s (avg 3m26s)