Dummy data should be time-invariant

[evansd]

Dummy data generation is currently deterministic in one sense as we used a fixed PRNG seed. However, when we generate dates we do so as offsets from today's date which means that, while the pattern of dates we generate is deterministic, the actual dates produced change every day:

ehrql/ehrql/dummy_data/generator.py

Lines 182 to 194 in f5b0d5f

	def generate_patient_facts(self, patient_id):
	# Seed the random generator using the patient_id so we always generate the same
	# data for the same patient
	self.rnd.seed(f"{self.random_seed}:{patient_id}")
	# TODO: We could obviously generate more realistic age distributions than this
	date_of_birth = self.today - timedelta(days=self.rnd.randrange(0, 120 * 365))
	age_days = self.rnd.randrange(105 * 365)
	date_of_death = date_of_birth + timedelta(days=age_days)
	self.date_of_birth = date_of_birth
	self.date_of_death = date_of_death if date_of_death < self.today else None
	self.events_start = self.date_of_birth
	self.events_end = min(self.today, date_of_death)

This was done as the quickest way of generating "reasonable" looking dates but we've always known its inadequate:

ehrql/ehrql/dummy_data/generator.py

Lines 42 to 45 in f5b0d5f

	# TODO: I dislike using today's date as part of the data generation because it
	# makes the results non-deterministic. However until we're able to infer a
	# suitable time range by inspecting the query, this will have to do.
	self.today = today if today is not None else date.today()

Ideally, we'd inspect all the various date filters used in a given dataset definition and then, based on that, come up with a suitable date range for generating dummy events.