Remove _pipeline from linker and refactor CTE pipeline

[RobinL]

This PR:

• eliminates the shared linker._pipeline, so that all SQL operations create and use fresh pipeline(s).
• eliminated associated linker methods linker._enqueue_sql. linker._execute_sql_pipeline
• all SQL execution now happens via a pipeline that gets passed to database_api.sql_pipeline_to_splink_dataframe
• eliminates confusing linker methods _initialise_df_concat and _initialise_df_concat_with_tf, instead relying on more explicit calculations using functions in vertically_concatenate.py
• clarifies the naming of pipeline classes to be more precise (Task is now CTE, and SQLPipeline is now CTEPipeline)
• removes the option of passing input_dataframes to sql_pipeline_to_splink_dataframe, instead, input dataframes can be added to a CTEPipeline with .append_input_dataframe. Users can therefore add dataframes in places that make the logic flow more clearly
• Pipelines are now 'single use'. Once executed, they will not be allowed to be re-used; a fresh one has to be created. A pipeline.spent property enforces this.

Motivation for this PR

Consider the existing _initialise_df_concat_with_tf.

The return type and the mutations of state it performs are confusing:

• If you set materialise=True it returns a Splink dataframe.
• If you set materialise=False, it enqueues sql on linker._pipeline() and returns None

So it relies on:

• the ability to mutate linker._pipeline
• The ability to execute linker._pipeline, and reuse it

This function is not really compatible with the idea of using a fresh SQL pipeline each time we want to queue sql. You'd need to pass a pipeline in, but it's not clear what comes out.

By allowing input tables to be queued directly onto the CTEPipeline, then we can write a new function linker._enqueue_df_concat_with_tf which:

• If you set materialise=True, runs the SQL, and returns a CTEPipeline with the result already enqueued
• If you set materialise=False, enqueues the SQL without running it to the pipeline, and returns a CTEPipeine

Thus allowing us to replace all use of _initialise_df_concat_with_tf

Also closes #1696

Reviewing

The main changes have been on:

• pipeline.py
• linker.py
• database_api.py

All the other changes are all just downstream consequences of changing those files

I was getting weirdness with the tests so you'll see i had to reset the cache

Related possible future PRs

• Should the pipeline have a pipeline.to_splink_dataframe(db_api) method? This feels clearer than having to call db_api.sql_pipeline_to_splink_dataframe(pipeline)

[ADBond]

I think this is great, looks like a big QoL improvement! Feels a lot clearer treating input frames in this way also - used to often get a bit tripped up by the old way, as it didn't line up between when you need/use them.

Haven't looked in detail at everything, but happy that the shape of this is good, and sure we can pick up any small issues if any arise