test(ci): ensure that we do not skip all of test_client.py if polars is not installed

[cpcloud]

Avoid skipping all of test_client.py when polars is not installed.

[cpcloud]

Confirmed that test_client.py is now running for Exasol: https://github.com/ibis-project/ibis/actions/runs/9973508235/job/27558993516?pr=9617#step:16:76

[cpcloud]

cc @jcrist for all things reference count related 😬

This assertion passes, but if you apply this diff (inlining the cached_table.to_pandas() call, then refcount assertion fails. Here's the diff:

diff --git a/ibis/backends/tests/test_expr_caching.py b/ibis/backends/tests/test_expr_caching.py
index 9551f4133..f38df40c4 100644
--- a/ibis/backends/tests/test_expr_caching.py
+++ b/ibis/backends/tests/test_expr_caching.py
@@ -73,8 +73,7 @@ def test_persist_expression_multiple_refs(backend, con, alltypes):
     op = non_cached_table.op()
     cached_table = non_cached_table.cache()
 
-    cached_df = cached_table.to_pandas()
-    backend.assert_frame_equal(non_cached_table.to_pandas(), cached_df)
+    backend.assert_frame_equal(non_cached_table.to_pandas(), cached_table.to_pandas())
     assert sys.getrefcount(op) == 2
 
     name = cached_table.op().name

Digging around a bit with gc.get_referrers, it looks like a stack frame from inside pyspark's pandas/conversion.py is holding a reference to cached_table in the inlined case but not when the DataFrame is lifted out into a variable.

Would appreciate any thoughts you have here. I'll probably keep the original code and xfail pyspark for now if there's no obvious way forward here.

[cpcloud]

I'm pushing some other commits, and I'll avoid force pushing for now. You'll want to poke around at 6d0971b.

[gforsyth]

Looks good to me.