Reduce parametric tests

[gsakkis]

The unit test suite has grown too large, mainly due to many parameter combinations generated by parametrize_for_dataset. This PR reduces the number of generated parametrizations by:

• merging some parameters (x_sparse/y_sparse -> sparse and similarly for all x_*/y_* parameters)
• removing parameters that don't offer much/any value from some tests (e.g. batch_size, shuffle_buffer_size, num_workers from test_unequal_num_keys).

While at it, some new tests were added:

• shuffle_buffer_size was tested by default only for a single value (16), now it is tested for both 0 and 16.
• test_csr was added in test_tensorflow (currently asserts that NotImplementedError is raised).

Number of tests:

• at the base of this branch (defb304):

$ pytest tests/ --collect-only | tail -1
======================== 10572 tests collected in 4.09s ========================

• at this branch (582acb0):

$ pytest tests/ --collect-only | tail -1
======================== 3336 tests collected in 3.25s =========================

[shortcut-integration]

This pull request has been linked to Shortcut Story #20800: Simplify parametric tests.

[ktsitsi]

LGTM

[georgeSkoumas]

LGTM