refactor(bigframes): Make join nullity optimizations more robust

[TrevorBergeron]

Thank you for opening a Pull Request! Before submitting your PR, there are a few things you can do to make sure it goes smoothly:

• Make sure to open an issue as a bug/issue before writing your code! That way we can discuss the change, evaluate designs, and agree on the general idea
• Ensure the tests and linter pass
• Code coverage does not decrease (if any source code was changed)
• Appropriate docs were updated (if necessary)

Fixes #<issue_number_goes_here> 🦕

[gemini-code-assist]

Code Review

This pull request introduces a nulls_equal attribute to JoinNode to explicitly handle pandas-style null equality in joins. It includes a new rewrite rule, simplify_join, which optimizes SQL generation by removing unnecessary nullity checks when join columns are non-nullable. The feedback suggests refining the simplify_join logic to be less restrictive by using an and condition for nullability checks, ensuring the optimization is applied more broadly. Additionally, it is recommended to reorder the compiler steps to ensure all join nodes are optimized and to remove an unused import in the new rewrite module.

[gemini-code-assist]

The use of or here is too restrictive and leads to a regression in SQL quality (unnecessary COALESCE operations) when only one side of a join condition is nullable. Since NULL == NULL can only occur if both sides are nullable, we can safely set nulls_equal=False as long as at least one side of every join condition is non-nullable. This aligns with the previous logic used in the joins_nulls property and would resolve the regression seen in test_compile_fromrange snapshots.

Suggested change

	if node.left_child.field_by_id[left_ref.id].nullable or node.right_child.field_by_id[right_ref.id].nullable:
	return node
	if node.left_child.field_by_id[left_ref.id].nullable and node.right_child.field_by_id[right_ref.id].nullable:
	return node

[gemini-code-assist]

For consistency with the sqlglot compiler and to ensure that any JoinNodes introduced during the lowering of unsupported operations are also optimized, simplify_join should be called after _replace_unsupported_ops.

Suggested change

	result_node = result_node.bottom_up(rewrites.simplify_join)
	result_node = _replace_unsupported_ops(result_node)
	result_node = _replace_unsupported_ops(result_node)
	result_node = result_node.bottom_up(rewrites.simplify_join)

[gemini-code-assist]

The itertools module is imported but not used in this file.