Remove backend registration system

[evansd]

At present, backends define a backend_id and register themselves in a global dictionary so that they can be referenced by this ID in environment variables.

However, this requires importing all the backends so they can register themselves. And this causes circular import problems e.g.

$ python -c 'from databuilder.query_engines.sqlite import SQLiteQueryEngine'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/dave/projects/ebmdatalab/databuilder/databuilder/query_engines/sqlite.py", line 5, in <module>
    from databuilder.query_engines.base_sql import BaseSQLQueryEngine
  File "/home/dave/projects/ebmdatalab/databuilder/databuilder/query_engines/base_sql.py", line 8, in <module>
    from databuilder.backends.base import DefaultBackend
  File "/home/dave/projects/ebmdatalab/databuilder/databuilder/backends/__init__.py", line 3, in <module>
    from .databricks import DatabricksBackend
  File "/home/dave/projects/ebmdatalab/databuilder/databuilder/backends/databricks.py", line 2, in <module>
    from ..query_engines.spark import SparkQueryEngine
  File "/home/dave/projects/ebmdatalab/databuilder/databuilder/query_engines/spark.py", line 4, in <module>
    from databuilder.query_engines.base_sql import BaseSQLQueryEngine
ImportError: cannot import name 'BaseSQLQueryEngine' from partially initialized module 'databuilder.query_engines.base_sql' (most likely due to a circular import) (/home/dave/projects/ebmdatalab/databuilder/databuilder/query_engines/base_sql.py)

We should get rid of the registration system and allow the environment to specify the backend by its dotted path e.g.

OPENSAFELY_BACKEND='databuilder.backends.tpp.TPPBackend'

As well as fixing the circular import problem and simplifying the code this naturally supports out-of-tree backends, which I think is desirable.