Versions after `2.28.0` fail to infer grouping decoders after a date is selected from a data structure

The below code throws this type error on the effected versions, and merely works as expected on 2.28.0:

`TypeError: Unable to deterministically encode '2021-11-02' of type '<class 'datetime.date'>', please provide a type hint for the input of 'GroupByKey' [while running 'Create/Map(decode)']`

import typing
from datetime import date
import apache_beam as beam
from apache_beam.testing.test_pipeline import TestPipeline

with TestPipeline() as pipeline:
	today = date.today()
	results = (
		pipeline
		| beam.Create([(1, { 'd': today }), (1, { 'd': today })])
		| beam.MapTuple(lambda i, d: (d['d'], i)) # <-- this step only requires output type hints on versions after 2.28.0, and only if the date is being "projected" from some other data structure
		| beam.CombinePerKey(sum) # <-- if this aggregation is removed, the pipeline also works without error
	)

	results | beam.Map(print)

This stackoverflow issue is having the same problem:
https://stackoverflow.com/questions/69409693/how-do-i-use-a-datetime-date-value-in-apache-beam-groupby

It's possible to fix the errors by registering a `DateCoder` and adding output type hints to the projection `MapTuple` step, but since this isn't necessary in other situations and versions, it seems this is a bug. Our production pipelines will need to add many of these tedious type hints in order to work properly, so we're effectively blocked from upgrading to the newest version.