Fix Basix bugs surfaced by Pyrefly

[jhale]

Stacked on #1068. Fixes the real bugs Pyrefly surfaced there, shrinking the baseline from 22 to 7.

• ufl.py: annotate get_tensor_product_representation's base return type (was inferred -> None, tripping a false bad-override on every real subclass); fix a loop variable (e) used after its for loop in create_enriched_element, which was unbound whenever map_type was already given.
• test_continuity.py: facets rows mixed ndarray and dict as plain lists, so unpacking unioned every variable's type across the row — switched to tuples.
• test_bernstein.py, test_mappings.py (×2), test_interpolation_between_elements.py: non-exhaustive if/elif chains left a variable possibly unbound; added an else: raise NotImplementedError.
• test_tensor_products.py: the assertion comparing values1/values2 sat after its for fs in factors loop instead of inside it, so only the last tensor factor was ever checked; moved it in — also unbound if factors was ever empty.
• test_version.py: import basix._basixcpp directly so it's a statically visible attribute.

The 7 remaining baselined diagnostics are matplotlib/sympy stub artifacts unrelated to Basix's own code (per Pyrefly, mplot3d's Axes return can be None; sympy's third-party stubs don't model Symbol arithmetic against numpy arrays).

Verified: pyrefly check --error-stale-baseline → 0 errors; ruff clean; full pytest coverage for every touched test file plus test_ufl_wrapper.py passes (957 passed, 0 failed).